Cowan BBS 2001

BEHAVIORAL AND BRAIN SCIENCES (2000) 24, 87185
Printed in the United States of America
The magical number 4 in short-term

memory: A reconsideration
of mental storage capacity
Nelson Cowan
Department of Psychological Sciences,
University of Missouri, Columbia, MO 65211
cowanN@missouri.edu www.missouri.edu/~psycowan
Abstract: Miller (1956) summarized evidence that people can remember about seven chunks in short-term memory (STM) tasks. How-
ever, that number was meant more as a rough estimate and a rhetorical device than as a real capacity limit. Others have since suggested
that there is a more precise capacity limit, but that it is only three to five chunks. The present target article brings together a wide vari-
ety of data on capacity limits suggesting that the smaller capacity limit is real. Capacity limits will be useful in analyses of information
processing only if the boundary conditions for observing them can be carefully described. Four basic conditions in which chunks can be
identified and capacity limits can accordingly be observed are: (1) when information overload limits chunks to individual stimulus items,
(2) when other steps are taken specifically to block the recoding of stimulus items into larger chunks, (3) in performance discontinuities
caused by the capacity limit, and (4) in various indirect effects of the capacity limit. Under these conditions, rehearsal and long-term
memory cannot be used to combine stimulus items into chunks of an unknown size; nor can storage mechanisms that are not capacity-
limited, such as sensory memory, allow the capacity-limited storage mechanism to be refilled during recall. A single, central capacity limit
averaging about four chunks is implicated along with other, noncapacity-limited sources. The pure STM capacity limit expressed in
chunks is distinguished from compound STM limits obtained when the number of separately held chunks is unclear. Reasons why pure
capacity estimates fall within a narrow range are discussed and a capacity limit for the focus of attention is proposed.
Keywords: attention; enumeration; information chunks; memory capacity; processing capacity; processing channels; serial recall; short-
term memory; storage capacity; verbal recall; working memory capacity
1. Introduction to the problem of mental Hulme 1995). One possible resolution is that the focus of
storage capacity attention is capacity-limited, whereas various supplemen-
tary storage mechanisms, which can persist temporarily
One of the central contributions of cognitive psychology has
without attention, are time-limited rather than capacity-
been to explore limitations in the human capacity to store
limited (Cowan 1988; 1995). Other investigators, however,
and process information. Although the distinction between
have long questioned whether temporary storage concepts
a limited-capacity primary memory and an unlimited-
are necessary at all, and have suggested that the rules of
capacity secondary memory was described by James (1890),
learning and memory could be identical in both the short
Millers (1956) theoretical review of a magical number
and long term (Crowder 1993; McGeoch 1932; Melton
seven, plus or minus two is probably the most seminal pa-
1963; Nairne 1992; Neath 1998).
per in the literature for investigations of limits in short-term
At present, the basis for believing that there is a time
memory (STM) storage capacity. It was, in fact, heralded as
one of the most influential Psychological Review papers
ever, in a 1994 centennial issue of the journal. Millers ref- Nelson Cowan (Ph.D. 1980, University of Wiscon-
erence to a magical number, however, was probably a sinMadison) is Professor in the Department of Psy-
rhetorical device. A more central focus of his article was the chological Sciences at the University of Missouri
ability to increase the effective storage capacity through the Columbia. He has written one book (Cowan, N. 1995.
use of intelligent grouping or chunking of items. He ulti- Attention and memory: An integrated framework. Ox-
mately suggested that the specific limit of seven probably ford University Press) and edited another (1997. The
emerged as a coincidence. development of memory in childhood. Psychology Press),
Over 40 years later, we are still uncertain as to the nature and has 100 other publications on working memory, and
its relation to attention. He is former Associate Editor
of storage capacity limits. According to some current theo- of the Journal of Experimental Psychology: Learning,
ries there is no limit in storage capacity per se, but a limit Memory, and Cognition (19951999) and current As-
in the duration for which an item can remain active in STM sociate Editor of the Quarterly Journal of Experimental
without rehearsal (e.g., Baddeley 1986; Richman et al. Psychology (section A). He won the 1998 University of
1995). This has led to a debate about whether the limitation Missouri Chancellors Award for Research and Creative
is a magic number or magic spell (Schweickert & Boruff Activities.
1986) or whether rehearsal really plays a role (Brown &
2000 Cambridge University Press 0140-525X/00 $12.50 87

Cowan: The magical number 4
limit to STM is controversial and unsettled (Cowan et al. pacity limits but they are in line with Millers 7 6 2 (e.g.,
1997a; 1997b; Crowder 1993; Neath & Nairne 1995; Ser- still taken at face value by Lisman & Idiart 1995). (View 2)
vice 1998). The question is nearly intractable because any Short-term memory is limited by the amount of time that
putative effect of the passage of time on memory for a par- has elapsed rather than by the number of items that can be
ticular stimulus could instead be explained by a combination held simultaneously (e.g., Baddeley 1986). (View 3) There
of various types of proactive and retroactive interference is no special short-term memory faculty at all; all memory
from other stimuli. In any particular situation, what looks results obey the same rules of mutual interference, distinc-
like decay could instead be displacement of items from a tiveness, and so on (e.g., Crowder 1993). (View 4) There
limited-capacity store over time. If the general question of may be no capacity limits per se but only constraints such
whether there is a specialized STM mechanism is to be an- as scheduling conflicts in performance and strategies for
swered in the near future then, given the apparent unre- dealing with them (e.g., Meyer & Kieras 1997). (View 5)
solvability of the decay issue, general STM questions seem There are multiple, separate capacity limits for different
more likely to hinge on evidence for or against a chunk- types of material (e.g., Wickens 1984). (View 6) There are
based capacity limit. separate capacity limits for storage versus processing
The evidence regarding this capacity limit also has been (Daneman & Carpenter 1980; Halford et al. 1998). (View
controversial. According to one view (Wickens 1984) there 7) Capacity limits exist, but they are completely task-spe-
is not a single capacity limit, but several specialized capac- cific, with no way to extract a general estimate. (This may
ity limits. Meyer and Kieras (1997) questioned the need for be the default view today.) Even among those who agree
capacity limits to explain cognitive task performance; they with the 4-chunk thesis, moreover, a remaining possible
instead proposed that performance scheduling concerns ground of contention concerns whether all of the various
(and the need to carry out tasks in the required order) ac- phenomena that I will discuss are legitimate examples of
count for apparent capacity limits. The goal of this target ar- this capacity limit.
ticle is to provide a coherent account of the evidence to date These seven competing views will be re-evaluated in sec-
on storage capacity limits . tion 4 (sects. 4.3.14.3.7). The importance of identifying
One reason why a resolution may be needed is that, as the chunk limit in capacity is not only to know what that
mentioned above, the theoretical manifesto announcing limit is, but more fundamentally to know whether there is
the existence of a capacity limit (Miller 1956) did so with such a limit at all. Without evidence that a consistent limit
considerable ambivalence toward the hypothesis. Although exists, the concepts of chunking and capacity limits are
Millers ambivalence was, at the time, a sophisticated and themselves open to question.
cautious response to available evidence, a wealth of subse-
quent information suggests that there is a relatively con-
1.1. Pure capacity-based and compound STM estimates
stant limit in the number of items that can be stored in a
wide variety of tasks; but that limit is only three to five I will call the maximum number of chunks that can be re-
items as the population average. Henderson (1972, p. 486) called in a particular situation the memory storage capac-
cited various studies on the recall of spatial locations or of ity, and valid, empirically obtained estimates of this num-
items in those locations, conducted by Sperling (1960), ber of chunks will be called estimates of capacity-based
Sanders (1968), Posner (1969), and Scarborough (1971), to STM. Although that chunk limit presumably always exists,
make the point that there is a new magic number 4 6 1. it is sometimes not feasible to identify the chunks inasmuch
Broadbent (1975) proposed a similar limit of three items on as long-term memory information can be used to create
the basis of more varied sources of information including, larger chunks out of smaller ones (Miller 1956), and inas-
for example, studies showing that people form clusters of much as time- and interference-limited sources of infor-
no more than three or four items in recall. A similar limit in mation that are not strictly capacity-limited may be used
capacity was discussed, with various theoretical interpreta- along with capacity-limited storage to recall information. In
tions, by others such as Halford et al. (1988; 1998), Luck various situations, the amounts that can be recalled when
and Vogel (1997), and Schneider and Detweiler (1987). the chunks cannot be specified, or when the contribution
The capacity limit is open to considerable differences of of non-capacity-limited mechanisms cannot be assessed,
opinion and interpretation. The basis of the controversy will be termed compound STM estimates. These presum-
concerns the way in which empirical results should be ably play an important role in real-world tasks such as prob-
mapped onto theoretical constructs. Those who believe in lem-solving and comprehension (Daneman & Merikle
something like a 4-chunk limit acknowledge that it can be 1996; Logie et al. 1994; Toms et al. 1993). However, the the-
observed only in carefully constrained circumstances. In oretical understanding of STM can come only from knowl-
many other circumstances, processing strategies can in- edge of the basic mechanisms contributing to the com-
crease the amount that can be recalled. The limit can pre- pound estimates, including the underlying capacity limit.
sumably be predicted only after it is clear how to identify The challenge is to find sound grounds upon which to iden-
independent chunks of information. Thus, Broadbent tify the pure capacity-based limit as opposed to compound
(1975, p. 4) suggested that The traditional seven arises . . . STM limits.
from a particular opportunity provided in the memory span
task for the retrieval of information from different forms of
1.2. Specific conditions in which a pure storage
processing.
capacity limit can be observed
The evidence provides broad support for what can be in-
terpreted as a capacity limit of substantially fewer than It is proposed here that there are at least four ways in which
Millers 7 6 2 chunks; about four chunks on the average. pure capacity limits might be observed: (1) when there is an
Against this 4-chunk thesis, one can delineate at least seven information overload that limits chunks to individual stim-
commonly held opposing views: (View 1) There do exist ca- ulus items, (2) when other steps are taken specifically to
88 BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1

block the recoding of stimulus items into larger chunks, (3) cluded because, in the absence of pre-existing knowledge
when performance discontinuities caused by the capacity of the novel stimulus words, each word may be encoded as
limit are examined, and (4) when various indirect effects of multiple phonemic or syllabic subunits with only weak as-
the capacity limit are examined. Multiple procedures fit un- sociations between these subunits (resulting in an underes-
der each of these headings. For each of them, the central timate of capacity). As another example, sets of dots form-
assumption is that the procedure does not enable subjects ing familiar or symmetrical patterns would be excluded for
to group items into higher-order chunks. Moreover, the the opposite reason, that multiple dots could be perceived
items must be familiar units with no pre-existing associa- together as a larger object with non-negligible inter-dot as-
tions that could lead to the encoding of multi-object groups, sociations, so that each dot would not be a separate chunk
ensuring that each item is one chunk in memory. Such as- (resulting in an overestimate of capacity). It also is neces-
sumptions are strengthened by an observed consistency sary to exclude procedures in which the central capacitys
among results. contents can be recalled and the capacity then re-used (e.g.,
The first way to observe clearly limited-capacity storage if a visual array remains visible during recall) or, conversely,
is to overload the processing system at the time that the in which the information is not available long enough or
stimuli are presented, so that there is more information in clearly enough for the capacity to be filled even once (e.g.,
auxiliary or time-limited stores than the subject can re- brief presentation with a mask). In section 3, converging
hearse or encode before the time limit is up. This can be ac- types of evidence will be offered as to the absence of inter-
complished by presenting a large spatial array of stimuli item chunking in particular experimental procedures (e.g.,
(e.g., Sperling 1960) or by directing attention away from the a fixed number of items correctly recalled regardless of the
stimuli at the time of their presentation (Cowan et al. 1999). list or array size).
Such manipulations make it impossible during the presen- Finally, it is necessary to exclude procedures in which the
tation of stimuli to engage in rehearsal or form new chunks capacity limit must be shared between chunk storage and
(by combining items and by using long-term memory in- the storage of intermediate results of processing. One ex-
formation), so that the chunks to be transferred to the ample of this is the n-back task in which each item in a
limited-capacity store at the time of the test cue are the continuous series must be compared with the item that oc-
original items presented. curred n items ago (e.g., Cohen et al. 1997; Poulton 1954)
The second way is with experimental conditions de- or a related task in which the subject must listen to a series
signed to limit the long-term memory and rehearsal proof digits and detect three odd digits in a row (Jacoby et al.
cesses. For example, using the same items over and over on 1989). In these tasks, in order to identify a fixed set of the
each trial and requiring the recall of serial order limits sub- most recent n items in memory, the subject must continu-
jects ability to think of ways to memorize the stimuli ally update the target set in memory. This task requirement
(Cowan 1995); and rehearsal can be blocked through the may impose a heavy additional storage demand. These de-
requirement that the subject repeat a single word over and mands can explain why such tasks remain difficult even with
over during the stimulus presentation (Baddeley 1986). n 5 3.
The third way is to focus on abrupt changes or disconti- It may be instructive to consider a hypothetical version
nuities in basic indices of performance (proportion correct of the n-back task that would be taken to indicate the exis-
and reaction time) as a function of the number of chunks in tence of a special capacity limit. Suppose that the subjects
the stimulus. Performance on various tasks takes longer and task were to indicate, as rapidly as possible, if a particular
is more error prone when it involves a transfer of informa- item had been included in the stimulus set previously. Some
tion from time-limited buffers, or from long-term memory, items would be repeated in the set but other, novel items
to the capacity-limited store than when it relies on the con- also would be introduced. On positive trials, the mean re-
tents of capacity-limited storage directly. This results in action time should be much faster when the item had been
markedly less accurate and/or slower performance when presented within the most recent three or four items than
more than four items must be held than when fewer items when it was presented only earlier in the sequence. To my
must be held (e.g., in enumeration tasks such as that dis- knowledge, such a study has not been conducted. However,
cussed by Mandler & Shebo 1982). in line with the expectation, probed recall experiments have
Fourth, unlike the previous methods, which have in- resulted in shorter reaction times for the most recent few
volved an examination of the level of performance in the items (Corballis 1967).
memory task, there also are indirect effects of the limit in The present view is that a strong similarity in pure ca-
capacity. For example, lists of items tend to be grouped by pacity limits (to about 4 chunks on average) can be identi-
subjects into chunks of about four items for recall (Broad- fied across many test procedures meeting the above four
bent 1975; Graesser & Mandler 1978), and the semantic criteria. The subcategories of methods and some key refer-
priming of one word by another word or learning of con- ences are summarized in Table 1 (see sect. 3), and each area
tingencies between the words appears to be much more po- will be described in more detail in section 3 of the target ar-
tent if the prime and target are separated by about three or ticle.
fewer words (e.g., McKone 1995).
1.3. Definition of chunks
1.2.1. Other restrictions on the evidence. Although these
four methods can prevent subjects from amalgamating A chunk must be defined with respect to associations be-
stimuli into higher-order chunks, the resulting capacity es- tween concepts in long-term memory. I will define the term
timates can be valid only if the items themselves reflect in- chunk as a collection of concepts that have strong associa-
dividual chunks, with strong intra-chunk associations and tions to one another and much weaker associations to other
weak or (ideally) absent inter-chunk associations. For ex- chunks concurrently in use. (This definition is related to
ample, studies with nonsense words as stimuli must be ex- concepts discussed by Simon 1974.) It would be assumed
BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1 89

Table 1. Types of evidence of a capacity limit of about four items, with selected key references
(numbered according to the relevant section of the article)
3.1. Imposing an information overload

3.1.1. Visual whole report of spatial arrays (Sperling 1960)
3.1.2. Auditory whole report of spatiotemporal arrays (Darwin et al. 1972)
3.1.3. Whole report of unattended spoken lists (Cowan et al. 1999)
3.2. Preventing long-term memory recoding, passive storage, and rehearsal
3.2.1. Short-term, serial verbal retention with articulatory suppression (see Table 3 references; also Pollack et al. 1959; Waugh &
Norman 1965)
3.2.2. Short-term retention of unrehearsable material (Glanzer & Razel 1974; Jones et al. 1995; Simon 1974; Zhang & Simon 1985)
3.3. Examining performance discontinuities
3.3.1. Errorless performance in immediate recall (Broadbent 1975)
3.3.2. Enumeration reaction time (Mandler & Shebo 1982, Trick & Pylyshyn 1993)
3.3.3. Multi-object tracking (Pylyshyn et al. 1994)
3.3.4. Proactive interference in immediate memory (Halford et al. 1988; Wicklegren 1966)
3.4. Examining indirect effects of the limits
3.4.1. Chunk size in immediate recall (Chase & Simon 1973; Ericsson 1985; Ericsson et al. 1980; Ryan 1969; Wickelgren 1964)
3.4.2. Cluster size in long-term recall (Broadbent 1975; Graesser & Mandler 1978)
3.4.3. Positional uncertainty in recall (Nairne 1991)
3.4.4. Analysis of the recency effect in recall (Watkins 1974)
3.4.5. Sequential effects in implicit learning and memory (Cleeremans & McClelland 1991; McKone 1995)
3.4.6. Influence of capacity on properties of visual search (Fisher 1984)
3.4.7. Influence of capacity on mental addition reaction time (Logan 1988; Logan & Klapp 1991)
3.4.8. Mathematical modeling parameters (Halford et al. 1998; Kintsch & van Dijk 1978; Raaijmakers & Shiffrin 1981)
that the number of chunks can be estimated only when verbal recall, as groups of items recited together quickly,
inter-chunk associations are of no use in retrieval in the helpful because recall timing provides one good indication
assigned task. To use a well-worn example inspired by of chunking (see also Anderson & Matessa 1997). McLean
Miller (1956), suppose one tries to recall the series of let- and Gregg (p. 456) described three ways in which chunks
ters, fbicbsibmirs. Letter triads within this sequence can be formed: (a) Some stimuli may already form a unit
(FBI, CBS, IBM, and IRS) are well-known acronyms, and with which S is familiar. (b) External punctuation of the stim-
someone who notices that can use the information to assist uli may serve to create groupings of the individual elements.
recall. For someone who does notice, there are pre-existing (c) The S may monitor his own performance and impose
associations between letters in a triad that can be used to structure by selective attention, rehearsal, or other means.
assist recall of the 12-letter sequence. If we further assume The practical means to identify chunks directly is an im-
that there are no pre-existing associations between the portant issue, but one that is more relevant to future em-
acronyms, then the four of them have to occupy limited- pirical work than it is to the present theoretical review of al-
capacity storage separately to assist in recall. If that is the ready-conducted work, inasmuch as few researchers have
case, and if no other optional mnemonic strategies are in- attempted to measure chunks directly. Direct measures of
volved, then successful recall of the 12-item sequence indi- chunks can include empirical findings of item-to-item as-
cates that the pure capacity limit for the trial was at least sociations that vary widely between adjacent items in a list,
four chunks. (In practice, within the above example there being high within a chunk and low between chunks; item-
are likely to be associations between the acronyms. For ex- to-item response times that vary widely, being relatively
ample, FBI and IRS represent two U.S. government agen- short within a chunk and long between chunks; and sub-
cies, and CBS and IBM represent two large U.S. corpora- jective reports of grouping. For studies in which the main
tions. Such associations could assist recall. For the most dependent measure is not overt recall, measures of chunk-
accurate pure capacity-based limit, materials would have to ing for a trial must follow the trial immediately if it cannot
be selected so as to eliminate such special associations be- be derived from the main dependent measure itself. Schnei-
tween chunks.) Notice that the argument is not that long- der and Detweiler (1987, pp. 105106) provide an excel-
term memory fails to be involved in capacity-based esti- lent further discussion of how chunks can be identified
mates. Long-term memory is inevitably involved in through convergent measures.
memory tasks. The argument is that the purest capacity es- For most of the research that will be summarized in sec-
timates occur when long-term memory associations are as tion 3 below, however, the researchers provided no direct
strong as possible within identified chunks and absent be- evidence of chunking or its absence. The present assump-
tween those identified chunks. tion for these studies is that chunk size can be reasonably
If someone is given new material for immediate recall and inferred from the presence of the task demands described
can look at the material long enough before responding, new above in section 1.2, which should prevent inter-item
associations between the original chunks can be formed, re- chunking. The present thesis is that the great similarity of
sulting in larger chunks or, at least, conglomerates with empirically-based chunk limits derived using these guide-
nonzero associations between chunks. McLean and Gregg lines, reviewed in section 3, supports their validity because
(1967, p. 455) provided a helpful description of chunks in the guidelines yield a parsimonious, relatively uniform de-

scription of capacity limits of three to five chunks as the 50 semantic modules might exist, each specializing in a given
population average (with a maximum range of two to six class of words, e.g., for categories such as animals or vehicles.
chunks in individuals). Nevertheless, if the controller can remember only the four
most active buffers, the number of active semantic buffers
would be effectively only four buffers, regardless of the total
2. Theoretical framework number of modules. . . . Based on our interpretations of em-
pirical literature, the number of active semantic buffers seems
The most important theoretical point here is the identifica- to be in the range of three to four elements.
tion of conditions under which a capacity limit can be ob-
served (see sect. 1); reasons for this limit also are proposed. The present analysis, based on Cowan (1988; 1995), ba-
The theoretical model in this section provides a logical way sically agrees with Schneider and Detweiler, though with
to understand the empirical results presented in section 3. some differences in detail. First, it should be specified that
A fuller analytic treatment, consideration of unresolved is- the elements limited to four are chunks. (Schneider & Det-
sues, and comparison with other approaches is provided in weiler probably agreed with this, though it was unclear
section 4. from what was written.) Second, the justification for the
The basic assumptions of the present theoretical frame- particular modules selected by Schneider and Detweiler
work are (1) that the focus of attention is capacity-limited, (or by others, such as Baddeley,1986) is dubious. One can
(2) that the limit in this focus averages about four chunks in always provide examples of stimuli that do not fit neatly into
normal adult humans, (3) that no other mental faculties are the modules (e.g., spatial information conveyed through
capacity-limited, although some are limited by time and acoustic stimulation). Cowan (1988; 1995) preferred to
susceptibility to interference, and (4) that any information leave open the taxonomy, partly because it is unknown and
that is deliberately recalled, whether from a recent stimu- partly because there may in fact not be discrete, separate
lus or from long-term memory, is restricted to this limit in memory buffers. Instead, there could be the activation of
the focus of attention. This last assumption depends on the multiple types of memory code for any particular stimulus,
related premise, from Baars (1988) and Cowan (1988; with myriad possible codes. The same general principles of
1995), that only the information in the focus of attention is activation and de-activation might apply across all types of
available to conscious awareness and report. The identifi- code (e.g., the principle that interference with memory for
cation of the focus of attention as the locus of the capacity an item comes from the activation of representations for
limit stems largely from a wide variety of research indicat- other items with similar memory codes), making the iden-
ing that people cannot optimally perceive or recall multiple tification of particular discrete buffers situation-specific
stimulus channels at the same time (e.g., Broadbent 1958; and therefore arbitrary. Third, Cowan (1995) suggested
Cowan 1995), although most of that research does not pro- that the focus of attention and its neural substrate differ
vide estimates of the number of chunks from each channel subtly from the controller and its neural substrate, though
that occupy the focus of attention at any moment. There is they usually work closely together. In particular, for reasons
an additional notion that the focus of attention serves as a beyond the scope of this target article, it would be expected
global workspace for cognition, as described, for example, that certain types of frontal lobe damage can impair the
by Cowan (1995, p. 203) as follows: controller without much changing the capacity of the focus
of attention, whereas certain types of parietal lobe damage
Attention clearly can be divided among channels, but under the would change characteristics of the focus of attention with-
assumption of the unity of conscious awareness, the perceived
contents of the attended channels should be somehow inte-
out much changing the controller (see Cowan 1995). In the
grated or combined. As a simple supporting example, if one is present analysis, it is assumed that the capacity limit occurs
instructed to divide attention between visual and auditory within the focus of attention, though the control mecha-
channels, and one perceives the printed word dog and the nism is limited to the information provided by that focus.
spoken word cat, there should be no difficulty in determining In the next section, so as to keep the theoretical frame-
that the two words are semantically related; stimuli that can be work separate from the discussion of empirical evidence, I
consciously perceived simultaneously can be compared to one will continue to refer to evidence for a capacity-limited
another, as awareness serves as a global workspace. (Baars STM without reiterating that it is the focus of attention
1988) that presumably serves as the basis of this capacity limit.
Cowan (1995) also suggested two other processing lim- (Other, non-capacity-limited STM mechanisms that may be
its. Information in a temporarily heightened state of activa- time-limited contribute to compound STM measures but
tion, yet not in the current focus of attention, was said to be not to capacity-limited STM.) Given the usual strong dis-
time-limited. Also, the transfer of this activated information tinction between attention and memory (e.g., the absence
into the focus of attention was said to be rate-limited. Im- of memory in the central executive mechanism as discussed
portant to note, however, only the focus of attention was as- by Baddeley 1986), the suggested equivalence of the focus
sumed to be capacity-limited. This assumption differs from of attention and the capacity-limited portion of STM may
approaches in which there are assumed to be multiple ca- require some getting used to by many readers. With use of
pacity limits (e.g., Wickens 1984) or perhaps no capacity the term capacity-limited STM, the conclusions about ca-
limit (Meyer & Kieras 1997). pacity limits could still hold even if it were found that the
The assignment of the capacity limit to the focus of at- focus of attention is not, after all, the basis of the capacity
tention has parallels in previous work. Schneider and Det- limit.
weiler (1987) proposed a model with multiple storage A further understanding of the premise that the focus of
buffers (visual, auditory, speech, lexical, semantic, motor, attention is limited to about four chunks requires a discus-
mood, and context) and a central control module. They sion of working assumptions including memory retrieval,
then suggested (p. 80) that the control module limited the the role of long-term memory, memory activation, main-
memory that could be used: tenance rehearsal, other mnemonic strategies, scene co-

herence, and hierarchical shifting of attention. These are the term template to refer to large patterns of informa-
discussed in the remainder of section 2. In section 3, cate- tion that an expert can retain as a single complex chunk
gories of evidence will be explained in detail. Finally, in sec- (concerning expert information in long-term memory, see
tion 4, on the basis of the evidence, the theoretical view will also Richman et al. 1995).
be developed and evaluated more extensively with particu- The role of long-term memory is important to keep in
lar attention to possible reasons for the capacity limits. mind in understanding the size of chunks. When chunks are
formed in the stimulus field on the basis of long-term mem-
ory information, there should be no limit to the number of
2.1. Memory retrieval
stimulus elements that can make up a chunk. However, if
It is assumed here that explicit, deliberate memory retrieval chunks are formed rapidly through new associations that
within a psychological task (e.g., recall or recognition) re- did not exist before the stimuli were presented (another
quires that the retrieved chunk reside in the focus of atten- mechanism suggested by McLean & Gregg 1967), then it is
tion at the time immediately preceding the response. The expected that the chunk size will be limited to about four
basis of this assumption is considerable evidence, beyond items because all of the items (or old chunks) that will be
the scope of this article, that explicit memory in direct grouped to form a new, larger chunk must be held in the fo-
memory tasks such as recognition and recall requires at- cus of attention at the same time in order for the new intra-
tention to the stimuli at encoding and retrieval, a require- chunk associations to be formed (cf. Baars 1988; Cowan
ment that does not apply to implicit memory as expressed 1995). This assumption is meant to account for data on lim-
in indirect memory tasks such as priming and word frag- itations in the number of items per group in recall (e.g., see
ment completion (for a review, see Cowan 1995). There- section 2.7). It should be possible theoretically to increase
fore, any information that is deliberately recalled, whether existing chunk sizes endlessly, little by little, because each
it is information from a recent stimulus or from long-term old chunk occupies only one slot in the capacity-limited
memory, is subject to the capacity limit of the focus of at- store regardless of its size.
tention. In most cases within a memory test, information
must be recalled from both the stimulus and long-term
2.3. Memory activation
memory in order for the appropriate units to be entered
into the focus of attention. For example, if we attempt to It is assumed that there is some part of the long-term mem-
repeat a sentence, we do not repeat the acoustic waveform; ory system that is not presently in the focus of attention but
we determine the known units that correspond to what was is temporarily more accessible to the focus than it ordinar-
said and then attempt to produce those units, subject to the ily would be, and can easily be retrieved into that focus if it
capacity limit. is needed for successful recall (Cowan 1988; 1995). This ac-
A key question about retrieval in a particular circum- cessible information supplements the pure capacity limit
stance is whether anything about the retrieval process and therefore must be understood if we are to determine
makes it impossible to obtain a pure capacity-based STM that pure capacity limit.
estimate. A compound STM estimate can result instead if According to Baddeley (1986) and Cowan (1995), when
there is a source of information that is temporarily in a information is activated (by presentation of that informa-
highly accessible state, yet outside of the focus of attention. tion or an associate of it) it stays activated automatically for
This is particularly true when a subjects task involves the a short period of time (e.g., 2 to 30 sec), decaying from ac-
reporting of chunks one at a time, as in most recall tasks. In tivation unless it is reactivated during that period through
such a situation, if another mental source is available, the additional, related stimulus presentations or thought pro-
subject does not need to hold all of the to-be-reported in- cesses. In Baddeleys account, this temporary activation is
formation in the focus of attention at one time. In a trivial in the form of the phonological buffer or the visuospatial
example, a compound, supplemented digit capacity limit sketch pad. As mentioned above, there is some question
can be observed if the subject is trained to use his or her about the evidence for the existence of that activation-and-
fingers to hold some of the information during the task decay mechanism. Even if it does not exist, however, there
(Reisberg et al. 1984). The same is true if there is some in- is another route to temporary memory accessibility, de-
ternal resource that can be used to supplement the focus of scribed by Cowan et al. (1995) as virtual short-term mem-
attention. ory and by Ericsson and Kintsch (1995), in more theoret-
ical detail, as long-term working memory. For the sake of
simplicity, this process also will be referred to as activation.
2.2. The role of long-term memory
Essentially, an item can be tagged in long-term memory as
Whereas some early notions of chunks may have conceived relevant to the current context. For example, the names of
of them as existing purely in STM, the assumption here is fruits might be easier to retrieve from memory when one is
that chunks are formed with the help of associations in long- standing in a grocery store than when one is standing in a
term memory, although new long-term memory associa- clothing store because different schemas are relevant and
tions can be formed as new chunks are constructed. It ap- different sets of concepts are tagged as relevant in memory.
pears that people build up data structures in long-term Analogously, if one is recalling a particular list of items, it
memory that allow a simple concept to evoke many associ- might be that a certain item from the list is out of the focus
ated facts or concepts in an organized manner (Ericsson & of attention at a particular point but nevertheless is tem-
Kintsch 1995). Therefore, chunks can be more than just a porarily more accessible than it was before the list was pre-
conglomeration of a few items from the stimulus. Gobet sented. For example, if one is buying groceries based on a
and Simon (1996; 1998) found that expert chess players dif- short list that was not written down, a fruit forgotten from
fer from other chess players not in the number of chunks the list might be retrieved with a process resembling the
but in the size of these chunks. They consequently invoked following stream of thought: I recall that there were three

fruits on the list and I already have gotten apples and ba- such as the fact that the 12-letter string given above could
nanas . . . what other fruit would I be likely to need? The be divided into four 3-letter acronyms. Elaborative re-
data structure in long-term memory then allows retrieval. hearsal involves an active search for meaningful associa-
One difference between this mechanism and the short- tions between items. For example, if the items fish, brick
term decay and reactivation mechanism is that it is limited were presented consecutively, one might form an image of
by contextual factors rather than by the passage of time. a dead fish on a brick, which could be retrieved as a single
If there is no such thing as time-based memory decay, the unit rather than two unconnected units. Recoding and elab-
alternative assumption is that long-term working memory orative rehearsal are not intended as mutually exclusive
underlies phenomena that have been attributed to the mechanisms, but slightly different emphases on how long-
phonological buffer and visuospatial sketchpad by Badde- term memory information can be of assistance in a task in
ley (1986). In the present article, the issue of whether short- which memory is required. These, then, are some of the
term decay and reactivation exists will not be addressed. In- main mechanisms causing compound STM limits to be pro-
stead, it is enough to establish that information can be made duced instead of pure capacity-based STM limits.
temporarily accessible (i.e., in present terms, active), by one
means or another and that this information is the main data
2.6. Scene coherence
base for the focus of attention to draw upon.
The postulation of a capacity of about four chunks appears
to be at odds with the earlier finding that one can compre-
2.4. Maintenance rehearsal
hend only one stream of information at a time (Broadbent
In maintenance rehearsal, one thinks of an item over and 1958; Cherry 1953) or the related, phenomenologically-
over and thereby keeps it accessible to the focus of atten- based observation that one can concentrate on only one
tion (Baddeley 1986; Cowan 1995). One way in which this event at a time. A resolution of this paradox was suggested
could occur, initially, is that the rehearsal could result in a by Mandler (1985, p. 68) as follows:
recirculation of information into the focus of attention, re- The organized (and limited) nature of consciousness is illus-
activating the information each time. According to Badde- trated by the fact that one is never conscious of some half dozen
ley (1986), the rehearsal loop soon becomes automatic totally unrelated things. In the local park I may be conscious of
enough so that there is no longer a need for attention. A four children playing hopscotch, or of children and parents in-
subject in a digit recall study might, according to this no- teracting, or of some people playing chess; but a conscious con-
tion, rehearse a sequence such as 2, 4, 3, 8, 5 while using tent of a child, a chess player, a father, and a carriage is unlikely
the focus of attention to accomplish other portions of the (unless of course they form their own meaningful scenario).
task, provided that the rehearsal loop contains no more than According to this concept, a coherent scene is formed in the
could be articulated in about 2 sec. In support of that no- focus of attention and that scene can have about four sepa-
tion of automatization, Guttentag (1984) used a secondary rate parts in awareness at any one moment. Although the
probe task to measure the allocation of attention and found parts are associated with a common higher-level node, they
that as children matured, less and less attention was de- would be considered separate provided that there are no
voted to rehearsal while it was going on. special associations between them that could make their re-
It appears from many studies of serial recall with rehearsal- call mutually dependent. For example, four spices might be
blocking or articulatory suppression tasks, in which a recalled from the spice category in a single retrieval (to the
meaningless item or short phrase is repeated over and over, exclusion of other spices), but salt and pepper are directly
that rehearsal is helpful to recall (for a review, see Badde- associated and so they could only count as a single chunk in
ley 1986). Maintenance rehearsal could increase the ob- the focus of attention.
served memory limit as follows. An individual might recall This assumption of a coherent scene has some interest-
an 8-item list by rehearsing, say, five of the items while holding implications for memory experiments that may not yet
ing the other three items in the focus of attention. There- have been conducted. Suppose that a subject is presented
fore, maintenance rehearsal must be prevented before pure with a red light, a spoken word, a picture, and a tone in rapid
capacity can be estimated accurately. succession. A combination of long-term memory and sen-
sory memory would allow fairly easy recognition of any of
these events, yet it is proposed that the events cannot eas-
2.5. Other mnemonic strategies
ily be in the focus of attention at the same time. One possi-
With the possible exception of maintenance rehearsal, ble consequence is that it should be very difficult to recall
other well-known mnemonic strategies presumably involve the serial order of these events because they were not con-
the use of long-term memory. In recoding, information is nected into a coherent scene. They can be recalled only by
transformed in a way that can allow improved associations. a shifting of attention from the sensory memory or the
For example, in remembering two lines of poetry that newly formed long-term memory representation of one
rhyme, an astute reader may articulate the words covertly item to the memory representation of the next item, which
so as to strengthen the temporary accessibility of a phono- does not result in a coherent scene and is not optimal for
logical or articulatory code in addition to whatever lexical serial recall.
code already was strong. This phonological code in turn al-
lows the rhyme association to assist retrieval of activated in-
2.7. Hierarchical shifting of attention
formation into the focus of attention. Another type of re-
coding is the gathering of items (i.e., chunks corresponding Attentional focus on one coherent scene does not in itself
to stimuli as intended by the experimenter) into larger explain how a complex sequence can be recalled. To un-
chunks than existed previously. This occurs when an indi- derstand that, one must take into account that the focus of
vidual becomes aware of the associations between items, attention can shift from one level of analysis to another.

McLean and Gregg (1967, p. 459) described a hierarchical lasted the brief stimulus array, and that items could be re-
organization of memory in a serial recall task with long lists called to the extent that the information could be extracted
of consonants: At the top level of the hierarchy are those from that preattentive store. On partial report trials, a tone
cueing features that allow S to get from one chunk to an- indicated which row of the array the subject should recall
other. At a lower level, within chunks, additional cues en- (in a written form), but on whole report trials the subject
able S to produce the integrated strings that become his was to try to recall the entire array (also in written form).
overt verbal responses. An example of hierarchical organi- The ability to report items in the array depended on the de-
zation was observed by Graesser and Mandler (1978) in a lay of the partial report cue. When the cue occurred very
long-term recall task. The assumption underlying this re- shortly after the array, most or all of the four items in the
search was that, like perceptual encoding, long-term recall cued row could be recalled, but that diminished as the cue
requires a limited-capacity store to operate. It was expected delay increased, presumably because the sensory store de-
according to this view that items would be recalled in bursts cayed before the subject knew which sensory information
as the limited-capacity store (the focus of attention) was to bring into the more limited, categorical store.
filled with information from long-term memory, recalled, By the time the cue was 1 sec later than the array, it was
and then filled and recalled again. Studies of the timing of of no value (i.e., performance reached an asymptotically
recall have indeed found that retrieval from long-term low level). Subjects then could remember about 1.3 of the
memory (e.g., recall of all the fruits one can think of) oc- cued items from a row of 4. It can be calculated that at that
curs in bursts of about five or fewer items (see Broadbent point the number of items still remembered was 1.3 3 3
1975; Mandler 1975). Graesser and Mandler (1978, study (the number of rows in the array) or about 4. That was also
2) had subjects name as many instances of a semantic cate- how many items subjects could recall on the average on tri-
gory as possible in 6 min. They used a mathematical func- als in the whole report condition, in which no partial re-
tion fit to cumulative number of items recalled to identify port cue was provided. The limit of four items was obtained
plateaus in the response times. These plateaus indicated in whole report across a large variety of arrays differing in
about four items per cluster. They also indicated, however, the number, arrangement, and composition of elements.
that there were lengthenings of the inter-cluster interval Thus, a reasonable hypothesis was that subjects could read
that defined superclusters. Presumably, the focus of atten- about four items out of sensory memory according to a
tion shifted back and forth between the supercluster level process in which the unlimited-capacity, fading sensory
(at which several subcategories of items are considered) store is used quickly to transfer some items to a limited-
and the cluster level (at which items of a certain subcate- capacity, categorical store (according to the present theo-
gory are recalled). An example would be the recall from retical framework, the focus of attention).
the fruit category as follows: applebananaorangepear One could illustrate the results of Sperling (1960) using
(some common fruits); grapesblueberriesstrawberries Figure 1, which depicts the interaction of nested faculties
(smaller common fruits); pineapplemango (exotic fruits); in the task in a manner similar to Cowan (1988; 1995).
watermeloncantaloupehoneydew (melons). By shifting Within that theoretical account, sensory memory is as-
the focus to higher and lower levels of organization it is pos- sumed to operate through the activation of features within
sible to recall many things from a scene. I assume that the long-term memory (an assumption that has been strength-
capacity limit applies only to items within a single level of ened through electrophysiological studies of the role of re-
analysis, reflecting simultaneous contents of the focus of at- activation in automatic sensory memory comparisons; see
tention. Cowan et al. 1993). The nesting relation implies that some,
but not all, sensory memory information also is in the focus
of attention at a particular moment in this task. In either the
3. Empirical evidence for the capacity limit whole report condition or the partial report condition, the
limited capacity store (i.e., the focus of attention) can be
3.1. Capacity limits estimated with information overload filled with as many of the items from sensory memory as the
One way in which long-term recoding or rehearsal can be limited capacity will allow; but in the partial report condi-
limited is through the use of stimuli that contain a large tion, most of these items come from the cued row in the ar-
number of elements for a brief period of time, overwhelm- ray. Because the display is transient and contains a large
ing the subjects ability to rehearse or recode before the ar- amount of information, subjects have little chance to in-
ray fades from the time-limited buffer stores. This has been crease the amount recalled through mnemonic strategies
accomplished in several ways. such as maintenance or elaborative rehearsal. Part A of the
figure represents whole report and shows that a subset of
3.1.1. Visual whole report of spatial arrays. One study the items can be transferred from activated memory to the
(Sperling 1960) will be explained in detail, as it was among capacity-limited store. Part B of the figure, represent-
the first to use the logic just described. It revealed evidence ing partial report, shows that the items transferred to the
for both (1) a brief, pre-attentive, sensory memory of un- capacity-limited store are now confined to the cued items
limited capacity and (2) a much more limited, post-attentive (filled circles), allowing a larger proportion of those items
form of storage for categorical information. Sperlings re- to be reported.
search was conducted to explore the former but it also was A word is in order about the intent of this simple model
informative about the latter, limited-capacity store. On shown in Figure 1. It is not meant to deny that there are im-
every trial, an array of characters (e.g., 3 rows with 4 letters portant differences between more detailed structures such
per row) was visually presented simultaneously, in a brief as the phonological store and the visuospatial sketch pad of
(usually 50-msec) flash. This was followed by a blank screen. Baddeley (1986). However, the model is meant to operate
It was assumed that subjects could not attend to so many on a taxonomically inclusive level of analysis. It seems likely
items in such a brief time but that sensory memory out- that there are other storage structures not included in Bad-

pacity limit in chunks, and it is that capacity limit that is of
primary concern here.
There are also a number of other theoretical suggestions
that are consistent with the present approach but with dif-
ferent terminology and assumptions. For example, the ap-
proach appears compatible with a model proposed recently
by Vogel et al. (1998). Their conceptual short-term mem-
ory would correspond to the activated portion of long-term
memory in the model of Cowan (1988), whereas their vi-
sual working memory would correspond to the focus of at-
tention. Potential differences between the approaches ap-
pear to be that what they call conceptual memory could,
according to Cowan (1988), include some physical features;
and what they call visual working memory would, accord-
ing to Cowan (1988), prove to be one instance of the focus
of attention, a central structure that represents conscious
information from all modalities. The most critical similarity
between the models for present purposes is that the capac-
ity limit shows up in only one place (the visual working
memory or focus of attention), not elsewhere in the model.
With these theoretical points in mind, we can return to a
consideration of Sperlings study. The observed limit to
about four items in whole report theoretically might be at-
tributed to output interference. However, studies by Pash-
ler (1988) and Luck and Vogel (1997), in which output in-
terference was limited, militate against that interpretation.
In one experiment conducted by Luck and Vogel, for ex-
ample, subjects saw an array of 1 to 12 small colored squares
for 100 msec and then, after a 900-msec blank interval, an-
other array that was the same or differed in the color of one
square. The subject was to indicate whether the array was
the same or different. Thus, only one response per trial was
required. Performance was nearly perfect for arrays of one
to three squares, slightly worse with four squares, and much
worse at larger array sizes. Very similar results were ob-
Figure 1. Illustration of the processing in (A) whole report pro- tained in another experiment in which a cue was provided
cedures and (B) partial report procedures according to the nested to indicate which square might have changed, reducing de-
processes framework suggested by Cowan (1988; 1995). In each cision requirements. Some of their other experiments clar-
figure, information is elevated from activated storage (jagged ify the nature of the item limit. The four-item limit was
lines) to the limited-capacity store, which is the focus of attention shown to apply to integrated objects, not features within ob-
(large circle), until the latter is full. Small circles represent items jects. For example, when objects in an array of bars could
in the array, and those with arrows enter the focus of attention. differ on four dimensions (size, orientation, color, and pres-
With partial report (B), the cued items (filled circles) preferen- ence or absence of a central gap), subjects could retain all
tially enter the limited-capacity focus of attention. four dimensions at once as easily as retaining any one. The
performance function of proportion correct across increas-
ing array size (i.e., increasing number of array items) was
deleys model, such as memory for nonverbal sounds and practically identical no matter how many stimulus attri-
for tactile stimuli. In principle, moreover, these separate butes had to be attended at once. This suggested that the
structures could share important functional properties capacity limit should be expressed in terms of the number
(e.g., incoming stimuli requiring a particular kind of coding of integrated objects, not the number of features within ob-
interfere with the short-term retention of existing repre- jects. The objects serve as the chunks here.
sentations using similar coding) and could operate based on Broadbent (1975) noted that the ability to recall items
similar neural mechanisms of activation. Given that we do from an array grows with the visual field duration: for the
not know the taxonomy of short-term stores, they are rep- first fiftieth of a second or so the rate of increase in recall is
resented together in the model according to their common extremely fast, and after that it becomes slower. He cites
principle, that they are activated portions of the long-term Sperlings (1967) argument that in the early period, items
memory system. This activated memory includes both are read in parallel into some visual store; but that, after it
physical features and conceptual features. What is critical fills up, additional items can be recalled only if some items
for the present purposes is that all of the storage structures are read (more slowly) into a different, perhaps articulatory
making up activated memory are assumed not to have ca- store. Viewed in this way, the visual store would have a ca-
pacity limitations. Instead, they are assumed to be limited pacity of three to five items, given that the performance
because of memory decay, interference from subsequent function rapidly increases for that number of items. How-
stimuli, and/or some other basis of temporary accessibility. ever, the visual store could be a central capacity limit (as-
Only the focus of attention is assumed to have a fixed ca- sumed here to be the focus of attention) rather than visu-

ally specific as the terminology used by Sperling seems to subjects ignored lists of digits presented through head-
imply. phones. Occasionally (just 16 times in a session), 1 sec af-
A related question is what happens when access to the ter the onset of the last spoken word in a list, the rhyming
sensory memory image is limited. Henderson (1972) pre- game disappeared from the screen and a memory response
sented 3 3 3 arrays of consonants, each followed by a mask- screen appeared shortly after that, at which time the sub-
ing pattern after 100, 400, 1,000, or 1,250 msec. This was ject was to use the keypad to report the digits in the spoken
followed by recall of the array. Although the number of con- list. Credit was given for each digit only if it appeared in the
sonants reported in the correct position depended on the correct serial position. Relative to a prior memory span task
duration of the array, the range of numbers was quite sim- result, lists were presented at span length and at lengths of
ilar to other studies, with means for phonologically dissim- span-1 (i.e., lists one item shorter than the longest list that
ilar sets of consonants ranging from about 3 with 100-msec was recalled in the span task), span-2, and span-3. A control
exposure times to about 5.5 with 1250-msec exposure condition in which subjects attended to the digits also was
times. This indicates that most of the transfer of informa- presented, before and after the ignored-speech session. In
tion from sensory storage to a limited-capacity store occurs the attended-speech control condition, the number of dig-
rather quickly. its recalled was higher than in the unattended condition,
A similar limit may apply in situations in which a scene is and it increased with list length. However, in the ignored-
changed in a substantial manner following a brief interrup- speech condition, the mean number of items recalled re-
tion and people often do not notice the change (e.g., Simons mained fixed at a lower level regardless of list length. The
& Levin 1998). Rensink et al. (1997) proposed that this limit level was about 3.5 items in adults, and fewer in children.
may occur because people can monitor only a few key ele- This pattern is reproduced in Figure 2. It is important that
ments in a scene at one time. the number correct remained fixed across list lengths in the
ignored-speech condition, just as the whole-report limit re-
3.1.2. Auditory whole report of spatiotemporal arrays. mained fixed across array sizes in Sperling (1960). It is this
Darwin et al. (1972) carried out an experiment that was pattern that is crucial for the conclusion that there is a fixed
modeled after Sperlings (1960) work, but with stimuli pre- capacity limit.
sented in the auditory modality. On each trial, subjects re- It is important also to consider individual-subject data.
ceived nine words in a spatiotemporal array, with sequences Sperlings (1960) data appeared to show that his very few,
of three spoken items (numbers and letters) presented over highly trained individuals had capacity limits in the range of
headphones at left, center, and right locations simultane- about 3.54.5. In the study of Cowan et al. (1999), results
ously for a total array size of nine items. The partial report from 35 adults are available even though only the first 24 of
cue was a visual mark indicating which spatial location to re- these were used in the published study. Figure 3 shows
call. The results were quite comparable to those of Sperling each adult subjects mean number correct in the unat-
(1960). Once more the partial report performance declined tended speech task, as well each subjects standard error
across cue delays until it was equivalent to the whole report and standard deviation across unattended speech trials. It
level of about four items though, in this experiment in the is clear from this figure that individuals did not fit within a
auditory modality, the decline took about 4 sec rather than very narrow window of scores; their individual estimates of
1 sec as in vision, and the last item in each sequence was re- capacity ranged from as low as about 2 to as high as almost
called better than the first two items. In both modalities, 6 in one participant. One might imagine that the higher es-
the whole report limit may suggest the limited capacity for timates in some individuals were owing to residual atten-
storage of item labels in a consciously accessed form. tion to the supposedly ignored spoken digits, but the results
do not support that suggestion. For example, consider the
3.1.3. Whole report of ignored (unattended) spoken lists. subject shown in Figure 3 who had the best memory for ig-
In all of the partial report studies, the measure of short- nored speech. If that subject attended to the spoken digits
term memory capacity depended upon the fact that there that were to be ignored, then the result should have been a
were too many simultaneously presented items for all of positive slope of memory across the four list lengths, simi-
them to be processed at once, so that the limited-capacity lar to the attended-speech condition shown in Figure 2. In
mechanism was filled with items quickly. If items were pre- fact, however, that subjects scores across four list lengths
sented slowly and one at a time, the subject would be able had a slope of 20.35. Across all of the adult subjects, the
to use mnemonic processes such as rehearsal (e.g., Badde- correlation between memory for ignored speech and slope
ley 1986) to expand the number of items that could be held, of the ignored speech memory function was r 5 2.19, n.s.
and therefore would be able to exceed the constraints of the The slight tendency was thus for subjects with better recall
limited-capacity store. If a way could be found to limit these to have less positive slopes than those with poorer recall.
mnemonic processes, it could allow us to examine pure ca- The slopes were quite close to zero (M 5 0.05, SD 5 0.32)
pacity in a test situation more similar to what is ordinarily and were distributed fairly symmetrically around 0. An-
used to examine STM (presumably yielding a compound other possible indication of attention to the supposedly ig-
STM estimate); namely immediate, serial verbal list recall. nored speech would be a tradeoff between memory and vi-
Cowan et al. (1999) limited the processing of digits in a sual task performance during the ignored speech session.
spoken list by having subjects ignore the items in the spo- However, such a tradeoff did not occur. The correlation be-
ken list until after their presentation. Subjects played a tween memory for unattended speech and reaction times
computer game in which the name of a picture at the cen- on the visual task was 2.33, n.s., the tendency being for sub-
ter of the screen was to be compared to the names of four jects with better memory for ignored speech also to display
surrounding pictures to indicate (with a mouse click) which slightly shorter reaction times on the visual task. The same
one rhymed with the central picture. A new set of pictures type of result was obtained for the relation between mem-
then appeared. As this visual game was played repeatedly, ory and visual task reaction times on a trial-by-trial basis

Figure 2. The number of items recalled in attended and unattended lists of digits. The flat pattern of recall of unattended digits as a
function of list length is taken as evidence of a capacity limit analogous to what Sperling (1960) reported. After Cowan et al. (1999).
within individuals. The mean within-subject correlation (1997), on visual storage capacity, resulted from individ-
was 2.08 (SD 5 .25), showing that the slight tendency was ual subject estimates of storage capacity ranging from 2.2
for a subjects trials that produced better memory to be ac- to 4.7, and a graduate student who spent months on the
companied by shorter mean reaction times on the preced- capacity-estimation tasks developed a capacity of about six
ing visual task. Thus, the memory capacity of up to six items items (Steven Luck, personal communication, January 18,
in certain individuals as measured in this technique and the 1999). These estimates are quite similar to the ones shown
individual differences in capacity seem real, not owing to in Figure 3 despite the great differences in procedures.
attention-assisted encoding. Figure 3 shows that individu- Similar estimates can be obtained from the study of Hen-
als standard errors (rectangles) were relatively small, and derson (1972), in which each consonant array was followed
that even the standard deviations (bars) of the best versus by a mask. For example, with a 400-msec field exposure du-
the worst rememberers did not overlap much. ration (long enough to access sensory memory once, but
The study of Cowan et al. (1999) is not the only one yield- probably not long enough for repeated access) and no sup-
ing individual difference information. For example, the plementary load, the six subjects mean number correct
data set reported as the first experiment of Luck and Vogel ranged from 3.0 to 5.1 items.
Figure 3. Number of items recalled in a memory for unattended speech task for each adult used in the procedure reported by Cowan
et al. (1999). Error bars depict trials within an individual.

Figure 4. Scatterplot of memory for unattended versus attended speech in each age group in the experiment of Cowan et al. (1999).
All of these results appear to require a modification of if it appeared in the correct serial position. Cowan et al. also
conclusions that could be drawn from the previous litera- examined results of a scoring procedure in which credit was
ture. In his ground-breaking review of memory span, given for any correct digit, regardless of the serial position.
Dempster (1981, p. 87) concluded that there is little or no Such results cannot be compared across list lengths because
evidence of either individual or developmental differences the probability of guessing correctly increases dramatically
in capacity. In the previous literature only processing with list length (given that each digit could occur only once
speeds were found related to span, but none of the previ- per list). Nevertheless, it is noteworthy that recall at all ages
ous developmental investigations examined memory with was more like a constant proportion correct across lists
strategic processing during reception of the list minimized lengths in this free scoring, not a constant number correct
so as to examine capacity. There do appear to be individual as in the serial position scoring. Adults and fourth-grade
and developmental differences in capacity. children were over 90% correct on lists of length 4 through
Figure 4 illustrates another intriguing point about Cowan 6, the lengths examined with this scoring procedure, and
et al. (1999). In this scatterplot of memory for unattended first-grade children were correct on 83, 80, and 83% of the
versus attended speech in individuals within each age lists at these three lengths. The item scoring raises the ques-
group, the equation line represents the case in which mem- tion of what it is that is held in a capacity-limited mech-
ory was equal in the two tasks. What the plot shows is that anism. It cannot simply be the items that are held, as the
memory was always better in the attended speech task, but free scoring does not show the limited-capacity pattern of
that the amount of improvement in the attended speech a constant number correct across list lengths. The digits
task relative to the unattended speech task was indepen- themselves may be stored in activated memory (e.g., audi-
dent of the level of performance on the unattended task. In tory sensory or phonological memory) and drawn from it
the attended condition the means (and SDs) were: for 35 into the focus of attention as needed. Instead, it might be
adults, 5.43 (0.78); for 26 fourth graders, 4.31 (0.88); and the mapping between the digits in memory and the serial
for 24 first graders, 3.48 (0.69). In the ignored speech con- positions in the list that would have to be held in capacity-
dition the comparable means were: for the adults, 3.51 limited storage.
(0.94); for the fourth graders, 2.99 (0.86); and for the first
graders, 2.34 (0.69). Notice that, among all groups, the ra-
3.2. Capacity limits estimated by blocking long-term
tio of mean attended to unattended numbers correct fell
memory recoding, passive storage, and rehearsal
within a narrow range, between 1.4 and 1.6. This pattern
suggests that attention at the time of reception of the list Verbal materials can be used under conditions that dis-
may add a process that is independent of the processes in- courage recoding and rehearsal, or materials that are in-
volved in memory for unattended speech. That process pre- trinsically difficult to recode, store, and rehearse can be
sumably is independent of the pure capacity limit and could used. These methods force subjects to rely primarily on
reflect the use of attention to form larger chunks. capacity-limited storage of chunks that were learned out of
It should be noted that the main scoring procedure used the laboratory or, at least, before the experimental trial in
by Cowan et al. (1999) credited correct recall of a digit only question.

3.2.1. Short-term, serial verbal retention with articulatory printed words, but not for spoken words (Baddeley et al.
suppression. The contribution of long-term memory can 1984). Finally, conditions with highly unusual stimulus pa-
be minimized by drawing the stimuli from the same, small rameters were eliminated. Unusually slow stimulus presen-
set on every trial and requiring the correct recall of serial tations (. 4 sec per word) were excluded because it might
order. Because the same items recur over and over, it is dif- be possible to insert rehearsals despite the articulatory sup-
ficult to retain long-term associations that help in the re- pression, as were unusually fast presentations (, 0.5 sec per
tention of serial order of the items on a particular trial. That word) because of encoding difficulty; and grouped presen-
is, in fact, the nature of the stimuli in most immediate, ser- tations were omitted because they encourage long-term re-
ial recall experiments that have been conducted. Further, coding of the list.
the contribution of rehearsal can be minimized by impos- Table 2 shows the results for all studies meeting these
ing articulatory suppression (Baddeley 1986; Murray 1968), constraints. I was able to find nine studies that included at
a secondary task in which the subject repeats, whispers, or least one experimental condition involving the immediate
mouths a rote utterance over and over during the presen- recall of spoken, monosyllabic words from a small set in the
tation of items (e.g., the, the, the . . . ) and sometimes presence of articulatory suppression. Among these studies
throughout recall itself if a nonspeech recall mode is used. I was able to derive 17 independent estimates of memory
Cowan et al. (1998) offered an account of what these vari- storage. There appears to be a striking degree of conver-
ables do when used together. They proposed that when new gence among the 17 estimates. All but one of the estimates
words are presented on every trial in a serial recall task, the fell within the range of three to five items, and most fell in
phonological portion of activated memory includes a the three- to four-item range. The only outlier was an esti-
phonological representation of the word sequence. How- mate of 2.4 items from Longoni et al. (1993). That low es-
ever, when the same words are used over and over on every timate is difficult to understand because the stimulus con-
trial, all of the representations of items from the memory ditions were almost identical to another experimental
set become active in memory, so that the memory items in condition in Longoni et al. that yielded an estimate of 3.4
the current list cannot necessarily be distinguished from items.
items used in previous trials. Rehearsal may allow a special The methods of estimation are described briefly in Table
representation of the to-be-recalled list to be constructed 2. The most commonly applicable method was to take the
in active memory even under these circumstances in which proportion correct at each serial position (or, when neces-
a small set of items is used over and over. Cowan et al. of- sary, an estimate of this proportion based on a figure) and
fered these assumptions to explain why articulatory sup- add the proportions across serial positions to arrive at the
pression has a much larger effect on performance for a number correct. In a probed recall experiment (e.g., Mur-
small set of words than for large sets of words that are not ray 1968) there is an initial list item for which the procedure
repeated from trial to trial. A small set of words used over produces no memory estimate; based on past research on
and over, along with articulatory suppression, may minimize primacy effects, the available proportion at this first serial
the contribution of articulatory and passive phonological position always was estimated at 0.8. For some studies, al-
storage factors in recall. It is only under these conditions, ternative assumptions led to alternative estimates of stor-
for example, that the word length effect, or advantage for age. For example, in the study of Peterson and Johnson
lists composed of short words, is eliminated (LaPointe & (1971), the dependent measure reported was the number
Engle 1990). Word length effects that remain even with ar- of lists recalled correctly, and to estimate items recalled one
ticulatory suppression when a large set of items is used can must make assumptions about the number of errors within
be explained on the grounds that phonological representa- the lists recalled incorrectly. Estimates of capacity are given
tions of these items are generated from long-term mem- in the table under a high assumption that at least four
ory (Besner 1987) and remain active despite articulatory items were recalled within each 5-item list, and under a
suppression. (An alternative interpretation of articulatory more moderate assumption that erroneous lists contained
suppression effects would state that suppression works by 1 or 2 errors (i.e., 4 or 3 correct items) equally often. It is
taking up processing capacity rather than by blocking re- the more moderate estimate that appears in the rightmost
hearsal. However, if that were true, suppression should im- column of the table. When the measure was memory span,
pair performance even when a large set of words is used. the estimate was taken as the span in conditions in which
Given that it does not, the alternative interpretation seems the articulatory suppression task can be presumed to have
wrong.) been most effective (e.g., in Cowan et al. 1987).
Before describing results of serial recall experiments Waugh and Norman (1965) impeded rehearsal in a dif-
with spoken stimuli and articulatory suppression, it is nec- ferent way, through instructions to the subjects not to re-
essary to restrict the admissible serial recall data in a few hearse. In their experiment, each list contained 16 spoken
other ways. Memory for multisyllabic words was excluded digits and the last digit was accompanied by a tone. It was
because these often might be retained as separate segments to serve as a probe, the same digit having occurred once be-
rather than integrated units (e.g., fire-man if morphemic fore somewhere in the list. The subject was to respond with
segments are used; um-brel-la if syllabic segments are the digit that had followed the probe digit when it was pre-
used). Memory for nonwords also was excluded because sented earlier, in the list. Results with an ordinary, 1-per-sec
one might retain them in terms of separate phonemic or syl- presentation rate (e.g., Waugh & Norman 1965, p. 91)
labic series even if they are monosyllabic. Only spoken showed that performance levels were much higher with 3
words were included because articulatory suppression or fewer items intervening between the target pair and the
seems to interfere with the retrieval of the phonological response (. .8) than it was with 4 or more intervening items
representation of printed words, but not of spoken words. (, .6). The transition between 3 and 4 intervening items
For example, articulatory suppression during the presenta- was abrupt. Note that with 3 intervening items in this task,
tion of a list eliminates phonological similarity effects for successful performance would require that the subjects

Table 2. Estimates of capacity from studies of immediate verbal memory with auditory presentation and articulatory suppression
Reference Data Source Method of Calculating Items in Storage Est.
Murray (1968) Figure 1, p. 682. Cued recall; auditory Add the proportions correct across probed serial
presentation with suppression. 6 letters. positions and assume recall of the first, unprobed 3.1
item in the list 5 .8. Thus, as List Length 6: .4 1
.2 1.4 1 .5 1 .8 5 2.3; 1 1 5 3.3.
same 7 letters same 3.2
Peterson & Johnson Table 2, p. 349 (5 letters, serial recall; 5 items, 45% of lists correct. High assumption
(1971) count during presentation, low- is that on the other 55% of trials, subjects get
similarity condition) Data 5 proportion 4 correct, for a mean of 5(.45) 1 4(.55) 5 4.45. 4.2
of lists recalled correctly. A more moderate assumption is 5(.45) 1 4(.28)
1 3(.27) 5 4.18.
Levy (1971) Table 1, p. 126. Neutral articulation 7 ser. posit. 3 .39 items/s.p. 5 2.73 items correct.
condition cued recall, simultaneous Add 0.8 items for the first, unprobed position 3.5
auditory presentation. 7 serial positions. 5 3.53 items.
same Table 2, p. 130. 9 serial positions 9 ser. posit. 3 .34 5 3.06, add 0.8 for first, 3.9
unprobed position 5 3.86 items.
Baddeley et al. (1975b) Figure 6, p. 585. Serial recall, mono- 5 serial positions 3 .7 items/s.p. correct 5 3.5 3.5
syllabic words, auditory presentation items.
with suppression.
Baddeley et al. (1984) Table 1, p. 236. Serial recall, mono- 5 serial positions 3 .64 items/s.p. correct 5 3.2 3.2
syllabic words, dissimilar items with items in each case.
suppression. Fast presentation
same Slower presentation Same 3.2
Cowan et al. (1987) Table 1, p. 514. Monosyllabic words, Span 5 estimate. (Omitted conditions in which
serial recall, span procedure, dis- articulatory suppression task was presumably
similar items with articulatory ineffective; whisper same letter once after each 4.0
suppression. Artic. Task: Whisper item, span 5 4.81; whisper same letter continu-
alphabet ously throughout study, span 5 4.86.)
same Task: Whisper next letter on each same 4.0
trial.
Longoni et al. (1993) Table 3, p. 17. Serial recall, distinct 6 serial positions 3 .57 correct 5 3.42 (Additional
items, suppression task = Whisper data from a very slow presentation rate of 5 sec
the. Presentation rate of 0.5 sec per item condition were omitted because rehearsal 3.4
per item. was possible; for that cond., 6 serial positions 3 .78
correct 5 4.68.)
same Table 4, p. 19. Whisper hiya 6 serial positions 3 .40 correct 5 2.40. It is not clear
during presentation and recall why such discrepant results obtained in these two 2.4
experiements.
Avons et al. (1994) Table 1, p. 215. Short words, 5 serial positions 3 .69 items/s.p. correct = 3.45 3.5
immediate recall (all had suppression). items.
Serial recall condition
same Probed recall condition. (Probed by the 5 serial positions 3 .72 items/s.p. 3.6
serial position of the item.) correct 5 3.60.
Hitch et al. (1996) Figure 4, p. 125. Auditory presentation Dependent measure 5 items correct 5 about
of items, suppression, recall in correct 4.0. (Omitted results for grouped lists, about 4.0
serial positions. 6.0).
memory extend back far enough to remember 4 items: the that was the most vivid for more recent items and faded
target pair and two intervening items. (The last intervening gradually across items.)
item was the probe, which did not have to be remembered.) Another way to limit rehearsal is to use a running mem-
Thus, this task leads to an estimate of four items in capacity- ory span procedure, in which a long list of items is pre-
limited short-term storage. (Performance levels with a very sented and the subject is unaware of the point at which the
fast, 4-per-sec presentation decreased rather more contin- test is to begin. Pollack et al. (1959) devised such a proce-
uously as a function of the number of intervening items, dure. In their Experiment 1, lists of 25, 30, 35, and 40 dig-
which possibly could reflect the heavier contribution of a its were presented. When the list ended, the task was to
time-limited source of activation, such as sensory memory, write down as many of the most recent items as possible,

making sure to write them in the correct serial positions at a time at different spatial locations on the computer
with respect to the end of the list. Under these conditions, screen. After a variable test delay, the response screen in-
the list was too long and continuous for rehearsal to do any cluded all of the dots and the task was to point to them in
good, and the obtained mean span was 4.2 digits. (Theo- the serial order in which they had been presented. There
retically, it might be possible for the subject continually to was very little loss of information over retention intervals of
compute, say, what were the last 5 items; but there is no task up to 30 seconds. The authors suggested that this stability
demand that would encourage such difficult work even if it of performance across test delays indicates that some sort
were feasible. The absence of on-line task requirements of rehearsal process was used. I would suggest that the so-
makes this task very different from the n-back tasks, which, called rehearsal process used here does not contaminate
as discussed earlier, do not meet the criterial for inclusion.) the estimate of storage because it is not a true rehearsal
It is possible to prevent rehearsal in yet another way, by process. Instead, it may be a process in which some of the
requiring processing between items rather than during the items, linked to serial position or order, are held in the ca-
presentation of items. Consider, for example, the working pacity-limited store. Each list presented by Jones et al. in-
memory span task of Daneman and Carpenter (1980) in cluded 4, 7, or 10 dots. It can be estimated from their pa-
which the subject must read sentences and also retain the per ( Jones et al., Fig. 2, p. 1011) that these three list lengths
final word of each sentence. The reading should severely led to means of 3.5, 3.8, and 3.2 items recalled in a trial, re-
limit rehearsal of the target words. Daneman and Carpen- spectively. These estimates were obtained by calculating
ter (1980, p. 455) reported a mean span of 3.15 words in this the mean proportion correct across serial positions and
circumstance. It is at first puzzling to think that subjects multiplying it by the number of serial positions.
could do this well, inasmuch as they might need some of the Several studies of the memory for unrehearsable mater-
capacity-limited storage space for processing the sentences ial produce estimates lower than 3.0. Glanzer and Razel
(unless storage and processing demands are totally separate (1974) examined the free recall of proverbs and estimated
as suggested by Daneman & Carpenter 1980 and by Hal- the short-term storage capacity using the method devel-
ford et al. 1998). Notice, however, that the word memory oped by Waugh and Norman (1965), based on the recency
load does not reach 3 until after the third sentence has been effect. The estimate was 2.0 proverbs in short-term storage
processed. This might well leave some of the limited stor- on the average. Glanzer and Razel also estimated the con-
age capacity available for sentence processing until the very tents of short-term storage for 32 different free recall ex-
end of the trial. periments, and found a modal value of 2.02.4 items in
storage, very comparable to what they found for the prov-
3.2.2. Short-term retention of unrehearsable material. A erbs. However, there is a potential problem with the Waugh
second way that time-limited stores can be eliminated from and Norman (1965) method of estimating the contents of
a measure of storage is with materials that, by their nature, short-term storage. They assumed that the most recent
cannot be rehearsed and thereby refreshed in active mem- items are recalled from either of two sources: short-term
ory. It is unlikely that items that cannot be rehearsed lend storage or long-term storage. The estimate of short-term
themselves easily to long-term recoding, either. An analysis storage is obtained by taking the list-medial performance
of one early study illustrates this distinction. Some verbal level to reflect long-term memory and assuming that the re-
materials are too long to be rehearsed (Baddeley 1986). Si- cency effect occurs because of this same memory plus the
mon (1974) examined this in an informal study using him- additional contribution of short-term memory. This as-
self as a subject, and tried to remember well-known ex- sumption is problematic, though, if the items in the recency
pressions such as four score and seven years ago, To be positions are not memorized in the same way but are more
or not to be, that is the question, and Alls fair in love and often recalled only with the short-term store and not with
war. He concluded that lists of three such phrases were the same contribution of long-term storage that is found for
all I could recall with reliability, although I could some- the earlier list items. This possibility is strengthened by the
times retain four. Of course, the number of words and syl- existence of negative recency effects in the final free recall
lables contained in these phrases was much larger. Else- of lists that previously had been seen in immediate recall
where in the article, for example, it was noted that seven (Craik et al. 1970). Glanzer and Razel consequently may
one-syllable words could be recalled. The present theoret- have overcorrected for the contribution of long-term mem-
ical assumption is that, in the recall of phrases, each phrase ory in the recency positions.
served as a previously learned chunk and also was too long Another low estimate was obtained for unrehearsable
to allow effective rehearsal; thus, by aiming the focus of at- material by Zhang and Simon (1985) using Chinese. In their
tention at the phrase level, four such phrases could be re- Experiment 1, the mean number of items recalled was 2.71
called despite their inclusion of many more units on a sub- when the items were radicals without familiar pronounce-
chunk level. In the recall of isolated words, in contrast, able names, and 6.38 (like the usual English memory span)
given that each word was much shorter than a phrase, it was when the items were characters with pronounceable, re-
presumably possible to use rehearsal to reactivate memory hearsable names, within which radicals were embedded. A
(and possibly to form new chunks larger than a single word) lower estimate for unrehearsable items is to be expected.
and therefore to increase the number of words recalled However, the fact that it was lower than three would not be
above what would be expected if each word were a separate expected if, as the authors asserted, there are over 200 such
chunk. This reasoning is supported by the fact that about radicals and educated Chinese people can recognize every
four unconnected spoken words can be recalled when re- radical (p. 194), making each radical a single visual chunk.
hearsal is blocked, as shown in Table 2. It seems possible that there are visual similarities among
Jones et al. (1995) carried out an experiment that reveals three or more radicals that tend to make them interfere
a capacity limit, though that was not the purpose of the ex- with one another in memory when radicals are presented in
periment. On each trial, a series of dots was presented one a meaningless series, preventing them from serving as in-

dependent chunks. Although this analysis is speculative, the of beans and threw them into a box, glancing at them briefly
basis of the discrepancy between these few estimates below and estimating their number, which was then counted for
3.0 and the estimates obtained in the many other experi- comparison. After over a thousand trials, he found that
ments taken to reflect a capacity limit (in the 35 chunk numbers up to four could be estimated perfectly, and up to
range) is an important area for future research. five with very few errors.
Kaufman et al. (1949) used the verb subitize to de-
scribe the way in which a few items apparently can be ap-
3.3. Capacity limits estimated with prehended and enumerated in a very rapid fashion (as if
performance discontinuities these items enter the focus of attention at the same time).
Although subjects in some procedures may be able to per- In contrast, when there are more items, the reaction time
form when there are more than four items, the function de- or the time necessary for accurate counting increases
scribing the quality or speed of performance sometimes steeply as the number of items increases (as if these items
shows a discontinuity when one reaches about four items must enter the focus of attention to be counted piecemeal,
(e.g., a much longer reaction time cost for each additional not all at once). Mandler and Shebo (1982) described the
item after the fourth item). Presumably, in these circum- history of the subitizing literature. As they note, subitizing
stances, some optional processing mechanism must be used has been observed via two main procedures: one in which
to supplement the capacity-limited store only if the stimuli the duration of an array is limited and the dependent mea-
exceed the capacity. This can occur in several ways as shown sure is the proportion of errors in estimating the number of
below. items in the array, and another method in which the array
stays on and the primary dependent measure is the reaction
3.3.1. Errorless performance in immediate recall. Broad- time to respond with the correct number. The results from
bent (1975) noted that we usually measure span as the num- the first of these methods seem particularly clear. For ex-
ber of items that can be recalled on 50% of the trials. How- ample, in results reported by Mandler and Shebo (1982, p.
ever, he cites evidence that the number of items that can be 8), the proportion of errors was near zero for arrays of 14
recalled reliably, with a very high accuracy, is about three or items (or for 13 items with a presentation duration as short
four and is much more resistant to modifications based on as 200 msec) and increased steeply after that, at a rate of
the nature of the items (Cardozo & Leopold 1963; see also about 15% per additional item until nearly 100% error was
Atkinson & Shiffrin 1968). That is, there is a flat perfor- reached with an array size of 11. The reaction time in-
mance function across list lengths until three or four items. creased slowly with array sizes of 13 and more steeply for
It stands to reason that when items beyond four are re- array sizes of 5 8. After that it leveled off (whereas it con-
membered, it is through the use of supplementary mne- tinued to increase at the same rate, for much higher array
monic strategies (such as rehearsal and chunking), not be- sizes, in procedures in which the array stayed on and the de-
cause of the basic storage capacity. pendent measure was the time to produce the number).
The average response was identical to the correct response
3.3.2. Enumeration reaction time. The ability to appre- for array sizes of 18, with an increasing degree of under-
hend a small number of items at one time in the conscious estimation as array size increased from 9 to 20. From the
mind can be distinguished from the need to attend to items present viewpoint, it would appear that three or four items
individually when a larger number of such items are pre- were subitized initially and about three or four more could
sented. This point is one of the earliest to be noted in psy- be added to the subitized amount without losing track of
chological commentaries on the limitations in capacity. which items had been counted.
Hamilton (1859) treated this topic at length and noted (vol. Alternative hypotheses about enumeration and related
1, p. 254) that two philosophers decided that six items could processes must be considered. Trick and Pylyshyn (1994a)
be apprehended at once, whereas at least one other (Abra- put forth a theory of subitizing suggesting that it is ca-
ham Tucker) decided that four items could be appre- pacity-limited (hence the limit to four items), but still not
hended. He went on to comment: The opinion [of six] ap- attention-demanding, and that it takes place at a point in
pears to me correct. You can easily make the experiment for processing intermediate between unlimited-capacity auto-
yourselves, but you must be aware of grouping the objects matic processes and serial or one-at-a-time attentive pro-
into classes. If you throw a handful of marbles on the floor, cesses. It was called the FINST (finger of instantiation) the-
you will find it difficult to view at once more than six, or ory in that there are a limited number of fingers of
seven at most, without confusion; but if you group them instantiation that can be used to define individual items in
into twos, or threes, or fives, you can comprehend as many the visual field. This theory is specific to vision, and it was
groups as you can units; because the mind considers these contrasted with a working memory theory in which subitiz-
groups only as units, it views them as wholes, and throws ing is said to occur because of a limit in the number of tem-
their parts out of consideration. You may perform the ex- porary memory slots.
periment also by an act of imagination. When the experi- The evidence used by Trick and Pylyshyn (1994a) to dis-
ment actually was conducted, however, it showed that tinguish between the theories is open to question. First, it
Hamiltons estimate was a bit high. Many studies have was shown that items could be subitized only if they were
shown that the time needed to count a cluster of dots or organized in a way that made them pop out of the sur-
other such small items rises very slowly as the number of roundings (the evidence of Trick & Pylyshyn 1993). Cer-
items increases from one to four, and rises at a much more tainly, this suggests that there is a pre-attentive stage of item
rapid rate after that. Jevons (1871) was probably the first ac- individuation, but perhaps the subitization occurs only af-
tual study, noting that Hamiltons conjecture was one of terward, contingent not only on this rapid item individua-
the very few points in psychology which can, as far as we yet tion as Trick and Pylyshyn said, but contingent also on the
see, be submitted to experiment. He picked up handfuls availability of slots. One reason to make this distinction is

that the phenomenon of popout clearly is not limited to four shyn & Storm 1988; Yantis 1992; for a recent review see Py-
items; it obviously occurs for much larger numbers of items. lyshyn et al. 1994). In the basic procedure, before the ob-
For example, when one looks inside a carton of eggs, all of jects move, some of them flash several times and then cease
the eggs appear to pop out against the surrounding carton. flashing. After that all of them wander randomly on the
It is the inclusion of individuated items in the enumeration screen and, when they stop, the subject is to report which
routine that is limited to about four. Another type of evi- dots had been flashing. The flavor of the results is described
dence used by Trick and Pylyshyn (1994a) was that there well by Yantis (1992, p. 307): Performance deteriorated as
was said to be no effect of a memory load on subitization, the number of elements to be tracked increased from 3 to
unlike counting. Logie and Baddeley (1987) were the main 5 [out of 10 on the screen]; tracking three elements was
authors cited in this regard, though subitization was not the viewed by most subjects as relatively easy, although not ef-
focus of their study. Logie and Baddeley did find that two fortless, while tracking 5 of 10 elements was universally
distractor tasks (articulatory suppression from repetition of judged to be difficult if not impossible by some subjects.
the word the, and tapping) had little effect in the subitiz- As in subitizing, one could use either FINST or working
ing range, whereas articulatory suppression had an effect in memory theories to account for this type of finding.
the counting range. However, these tasks can be carried out
relatively automatically and would not be expected to re- 3.3.4. Proactive interference in short-term memory. One
quire much working memory capacity (Baddeley 1986). can observe proactive interference (PI) in retrieval only if
For example, unlike counting backward as a distractor task, there are more than four items in a list to be retained (Hal-
which causes severe forgetting of a consonant trigram over ford et al. 1988). This presumably occurs because four or
an 18-sec distractor-filled period (Peterson & Peterson fewer items are, in a sense, already retrieved; they reside in
1959), articulatory suppression causes almost no loss over a a limited-capacity store, eliminating the retrieval step in
similar time period (Vallar & Baddeley 1982). Interference which PI arises. Halford et al. demonstrated this storage ca-
with articulatory processing can explain why articulatory pacity limit in a novel and elegant manner. They used vari-
suppression interfered with counting, for items over four in ant of Sternbergs (1966) memory search task, in which the
the array task and also for every list length within another subject receives a list of items and then a probe item and
task that involved enumeration of sequential events rather must indicate as quickly as possible whether the probe ap-
than simultaneous spatial arrays. The data of Logie and peared in the list. In their version of the task, modeled af-
Baddeley thus do seem to support the distinction between ter Wickens et al. (1981), lists came in sets of three, all of
subitizing and counting, but they do not necessarily support which were similar in semantic category (Experiment 1) or
the FINST theory over the working memory limitation the- rhyme category (Experiment 2). Thus, the first trial in each
ory of subitizing. set of three was a low-PI trial, whereas the last trial in the
Another type of evidence (from Trick & Pylyshyn 1994b) set was a high-PI trial. Experiment 1 showed that with lists
involved a cue validity paradigm (a variation of the proce- of 10 items, there were PI effects. With a list length of four,
dure developed by Posner et al. 1980). On each trial in most there was no PI. Experiment 2 showed that PI occurred for
of the experiments, two rectangles appeared; dots were to lists of six or more items, but not lists of four items. Pre-
appear in only one rectangle. The task was to count the dots. sumably, the items within a list of four did not have to be re-
Sometimes, there would be a cue (an arrow pointing to one trieved because they all could be present within a capacity-
rectangle or a flashing rectangle) to indicate slightly in ad- limited store at the same time. Also consistent with this sort
vance which rectangle probably would contain the dots. of interpretation, in 8- to 9-year-old children, PI was ob-
The cue was valid (giving correct information) on 80% of served with 4 items, but not 2 items in a list. The magnitude
the cued trials and invalid (giving incorrect information) on of growth of a capacity limit with age in childhood matches
20% of the cued trials. On other trials, no informative cue what was observed by Cowan et al. (1999) with a very dif-
was given. The validity of the cue affected performance in ferent procedure (see Fig. 2).
the counting range more than in the subitizing range, lead- In the Halford et al. (1988) study, it was the length of the
ing Trick and Pylyshyn (1994a) to view the results as sup- target list that was focused upon. We can learn more by ex-
portive of the FINST theory. However, there was still some amining also the effect of variations in the length of the list
effect of cue validity in the subitizing range, so the result is causing PI. Wickelgrens (1966) subjects copied a list of PI
less than definitive in comparing the FINST and working letters, a single letter to be recalled, and then a list of
memory accounts of subitizing. retroactive interference (RI) letters. The subject was to re-
Atkinson et al. (1976a) and then Simon and Vaishnavi call only the target letter. There were always eight letters in
(1996) investigated enumeration within afterimages so that one of the interference sets (the PI set for some subjects,
subjects would be unable to shift their gaze in a serial fash- the RI set for others), whereas the other interfering set
ion using eye movements. Both studies found that the could contain 0, 4, 8, or 16 letters. There was a large effect
subitizing limit remained at four items, with errors in enu- of the number of RI letters, with substantial differences be-
meration only above that number, even though subjects had tween any two RI list lengths. In contrast, when it was the
a long time to view each afterimage. Therefore, it seems PI set that varied in length, there was a difference between
that a focal attention strategy involving eye movements is 0 and 4 PI letters but very little effect of additional PI let-
important for visual enumeration of over four items, but not ters beyond 4. Wickelgren suggested that PI and RI both
at or below four items, the average number that subjects may generate associative interference, whereas RI additionally
be able to hold in the limited-capacity store at one time. generates another source of forgetting (either decay or stor-
age interference). Thus, associative interference would have
3.3.3. Multi-object tracking. Another, more recent line of been limited primarily to the four closest interfering items
research involves multi-object tracking of dots or small on either side of the target.
objects that move around on the computer screen (Pyly- A mechanism of PI in these situations can be suggested.

It seems likely that excellent, PI-resistant recall occurs The grouping limit seems to apply even for subjects who
when the active contents of the limited-capacity store are have learned how to repeat back strings of 80 or more dig-
to be recalled. When the desired information is no longer its (Ericsson 1985; Ericsson et al. 1980). These subjects did
active, the long-term memory record of the correct former so by learning to form meaningful chunks out of small
state of the limited-capacity store can be used as a cue to groups of digits, and then learning to group the chunks to-
the recall of the desired item(s). If several former limited- gether to form supergroups. At both the group and the
capacity states were similar in content, it may be difficult to supergroup levels, the capacity limit seems to apply, as de-
select the right one. Moreover, if the limited-capacity store scribed by Ericsson et al. (1980, p. 1182) for their first sub-
serves as a workspace in which items become associated ject who increased his digit span greatly: After all of this
with one another (Baars 1988; Cowan 1995), then it might practice, can we conclude that S.F. increased his short-term
be difficult to select the correct item from among several memory capacity? There are several reasons to think not . . .
present in the limited-capacity store simultaneously. The PI The size of S.F.s groups were almost always 3 and 4 digits,
results described above could then be interpreted as fol- and he never generated a mnemonic association for more
lows. For studies in Halford et al. (1988), target lists of more than 5 digits . . . He generally used three groups in his su-
than four items could not be held entirely within limited- pergroups and, after some initial difficulty with five groups,
capacity storage, so that a former state of the store had to never allowed more than four groups in a supergroup. Er-
be reconstituted. This could cause PI because some of the icsson (1985) reviewed details of the hierarchical grouping
target items may have shared a former limited-capacity structure in the increased-digit-span subjects and he re-
state with nearby items from a prior list, or because some viewed other studies of memory experts, which also re-
of the other former limited-capacity states would have con- vealed a similar grouping limit of 35 items. This limit to
tained items resembling the correct item. For subjects in the grouping process would make sense if the items or
Wickelgrens (1966) study who received a variable number groups to be further grouped together must reside in a
of PI letters, the target item would have been removed common, central workspace so that they can be linked. A
from the limited-capacity store by the presentation of eight limited-capacity store might be conceived in this way, as a
following RI letters. Therefore, at the time of recall, the workspace in which items are linked together (Baars 1988;
subject would have had to identify the former state of the Cowan 1995). The focus of attention, the presumed locus
limited-capacity store that contained the single target let- of limited-capacity storage, presumably must shift back and
ter. This same former state may also have included several forth from the super-group level to the group level.
of the adjacent letters, which could become confused with
the target letter. Only three or so letters adjacent to the tar- 3.4.2. Cluster size in long-term recall. As Shiffrin (1993)
get letter usually would have been in the limited-capacity pointed out, in a sense every cognitive task is a STM task
store at the same time as the target letter, and thus only because items must be active in a limited-capacity store at
those letters would contribute much to associative interfer- the time of recall, even though sometimes that can come
ence. In a broader context, this analysis may be one instance about only through the reactivation of long-term memories.
of a cue-overload theory of PI (cf. Glenberg & Swanson Assume, therefore, that it is necessary to represent items in
1996; Raaijmaker & Shiffrin 1981; Tehan & Humphreys the limited-capacity store to prepare them to be recalled.
1996; Watkins & Watkins 1975) asserting that recall is bet- That sort of mechanism should apply not only to immedi-
ter when fewer test items are associated with the cue used ate recall, but also to long-term recall. Bursts of responses
to recall the required information. should be observed as an individual fills the limited short-
term capacity, recalls the items held with that capacity, and
then refills it with more information retrieved from long-
3.4. Capacity limits estimated with indirect effects term storage.
So far we have discussed effects of the number of stimulus Broadbent (1975) used similar reasoning to motivate a
items on a performance measure directly related to the sub- study of grouping in long-term recall. He asked subjects to
jects task, in which recall of items in the focus of attention recall members of a learned category: the Seven Dwarfs,
is required. It is also possible to observe effects that are re- the seven colors of the rainbow, the countries of Europe, or
lated to the subjects task only indirectly by deriving a the- the names of regular television programs. There were some
oretical estimate of capacity from the presumed role of the important differences in the fluency of recall for these cat-
focus of attention in processing. egories. However, measurement of the timing of recall
showed bursts of 2, 3, or 4 items clustered together, and oc-
3.4.1. Chunk size in immediate recall. The magical num- casionally 5 items in a cluster. One of the 10 subjects pro-
ber 4 lurks in the background of the seminal article by duced a run of 6 rainbow colors, but otherwise the cluster
Miller (1956) on the magical number 7 6 2, which emphasizes were below 6. Thus, this study provides evidence for
sized the process of grouping elements together to form a short-term memory capacity limit even as applied to re-
larger meaningful units or chunks. The arrangement of call from a category in long-term memory. Wilkes (1975) re-
telephone numbers with groups of three and then four dig- viewed similar results in a more detailed study of pausing
its would not appear to be accidental, but rather an indica- within recall. Bower and Winzenz (1969, cited in Wilkes
tion of how many elements can be comfortably held in the 1975) showed that repetitions of a digit string within a
focus of attention at one time to allow the formation of a longer series resulted in improved recall over time only if
new chunk in long-term memory (Baars 1988; Cowan the grouping structure did not change each time the repe-
1995). Several investigators have shown that short-term tition occurred.
memory performance is best when items are grouped into Mandler (1967) suggested that the size of the limited-
sublists of no more than three or four (Broadbent 1975; capacity store is 5 6 2 items. He focused on a number of
Ryan 1969; Wickelgren 1964). experiments in which the items to be recalled could be di-

vided into a number of categories (e.g., fruits, clothing, ve- occurred (according to their criteria) most often without
hicles). Recall is superior when a list is recalled by cate- clustering and in clusters of two, three, four, or five in de-
gories; in a free recall task, subjects tend to recall clusters creasing order of frequency. In contrast, Graesser and Man-
organized by category even when the items were not or- dler (1978, p. 96) observed a limit in items per cluster hov-
dered by category in the stimulus list. Mandler suggested ering around 4.0 throughout a long recall protocol. A
that subjects left to their own devices typically could recall prediction of the present analysis is that if a particular pro-
5 6 2 categories and 5 6 2 items from each category that cedure results in a limit smaller than about four items, ex-
was recalled (e.g., apple, orange, banana, grape; shirt, tended practice with the task should eventually lead to a
pants, hat, and so on). (Mandler also relied on Tulving & plateau in performance at about four items per cluster.
Patkau 1962; and Tulving & Pearlstone 1966.)
The recall of separate items from a higher-level category 3.4.3. Positional uncertainty in recall. In a theoretical
in memory might be considered analogous to the retrieval mechanism discussed above and used earlier by Raaijmak-
of separate items from an array represented in sensory ers and Shiffrin (1981), items held simultaneously in the
memory. In the case of the sensory array, the items are re- limited-capacity store become associated with one another.
lated in that they all are part of a common visual field, but Additionally, their serial positions may become associated
are nevertheless separate perceptually (assuming they are with one another. In free recall, the associations can be
not organized into larger perceptual objects such as a clus- helpful because the thought of one item elicits the associ-
ter of rounded letters surrounded by angular letters). In ated items. However, in serial recall, the associations can
memory, the items are related in that they are associated to present a problem because it may be difficult to retrieve the
a common semantic category, but are nevertheless separate order of simultaneously held items. This type of account
conceptually (assuming they are not organized into clusters might explain positional uncertainty in serial recall. It pre-
such as salt-and-pepper within the spice category). The dicts that an item typically should be recalled no more than
clustering of items in recall presumably depends on the ab- three positions away from its correct position (assuming that
sence of an automatized routine for recall. Thus, one should sets of at most four items are present in limited-capacity
not expect clustering into groups of about four items for a storage at any moment). This prediction matches the data
task like recitation of the alphabet. The same applies to ex- well. For example, Nairne (1991) presented words aloud
tensive intra-category knowledge that can result in the re- with 2.5-sec onset-to-onset times between the words. Five
call of large chunks structures or templates (Gobet & Simon lists of 5 words each were followed by a 2-min distractor
1996; 1998). task and then a test in which the words were to be placed
When one focuses on flawless recall, the number is closer into their correct lists and locations within lists. The results
to three or four. Thus, Mandlers (1967, Fig. 7) summariza- showed that when an item was placed within the correct list,
tion of recall per category shows that when there were only its serial position was confused with at most three other se-
13 items in a category, these items were recalled flawlessly rial positions in the list. Nairne (1992) found a flattening of
(provided that the category was recalled at all). The num- the error functions with delayed testing. This change over
ber recalled within the category declined rather steadily time is compatible with increased difficulty of accessing
thereafter, from about 80% recalled from categories with information from a particular prior state of the limited-
46 items to about 20% recalled from categories of about capacity store, but with no evidence of a spread of uncer-
80 items. The growing absolute number of items recalled tainty to a larger range of serial positions. Within-list con-
from larger categories might be attributed to covert sub- fusions still occurred across about three items.
categorization or to long-term learning mechanisms, but
these will not allow the recall of all items. 3.4.4. Analysis of the recency effect in recall. Watkins
Why should similar constraints apply to the recall of (1974) reviewed research in which a long list of verbal items
items within a category and to the recall of categories? One was presented on each trial and the subject was to recall as
explanation is that the limited-capacity store can be used in many of those items as possible, without regard to the or-
more than one iteration. At one moment the categories are der of items. In such studies, recall is typically best for the
drawn into mind, and at the next moment the capacity is most recent items. This recency effect has been viewed as
consumed with items within the first category to be re- the result of the use of dual memory mechanisms, with a
called. An obvious question about such an account is how short-term memory mechanism used only for the last few
the categories are retained while the limited capacity is be- items (which typically are recalled first). Underlying this
ing used for items within a category. Presumably the im- view is the finding that the recency effect is quickly eroded
mediate consequence of the limited capacity store is to if a distractor-filled delay is imposed between the list and
form a better-organized long-term representation of the the recall cue (Glanzer & Cunitz 1966; Postman & Phillips
stimulus set. Thus, once a set of categories is brought into 1965), whereas the rest of the recall function is unaffected.
mind, these categories are combined into a long-term Several investigators have reasoned that it would be possi-
memory representation that can be accessed again as ble to estimate the contents of short-term memory by sub-
needed in the task. Ericsson and Kintsch (1995) provided a tracting out the contribution of long-term memory, but it is
detailed account of that sort of process. not clear exactly what assumptions one should make in or-
Finally, it should be noted that the capacity limit provides der to do so. Watkins ruled out some methods on the basis
only an upper bound for the clustering of items in recall. If of logical considerations, and compared the results of sev-
the rate of retrieval from memory is too slow, it might make eral favored methods (Tulving & Colotla 1970; Tulving &
sense for the individual to recall the items deposited in a Patterson 1968; Waugh & Norman 1965). Under a variety
capacity-limited store before that capacity has been used of test situations, these methods produced estimates of
up. Gruenewald and Lockhead (1980) obtained a pattern of short-term memory capacity ranging from 2.21 to 3.43
results in which the long-term recall of category exemplars (Watkins 1974, Table 1). For the method judged most ade-

quate (Tulving & Colotla 1970) the estimates ranged from together, indicated where the target would be on the sev-
2.93 to 3.35. enth, complex trial. Under these circumstances, subjects
One apparent difficulty for this interpretation of the re- succeeded in learning the contingencies and there was a
cency effect is that one can obtain it under filled test delays drop in performance when the contingencies were changed.
that are too protracted to permit the belief that a short-term However, as Stadler (1989) pointed out, any three of the
store is still in place. Bjork and Whitten (1974) presented a four critical trials were enough to determine the location of
series of printed word pairs for immediate free recall, with the target on Trial 7 in a set. If subjects remembered the
the pairs separated by silent intervals of 12 sec and with up outcomes of Trials 3, 4, and 6 and considered them to-
to 42 sec following the last item pair. These silent intervals gether, they theoretically could predict the outcome of Trial
were filled with a distracting arithmetic task to prevent re- 7. Given that subjects probably did not know which trials
hearsal. Under these conditions, a recency effect of 35 were predictive, they might only be able to use Trials 4, 5,
items emerged. The theoretical framework for understand- and 6, the last three trials, for prediction of Trial 7. These
ing these results, further developed by Glenberg and Swan- trials by themselves were predictive on a probabilistic,
son (1986), was one having to do with the ratio between the though not an absolute, basis. Thus, a limited-capacity store
inter-item interval and the retention interval, which could of four items could serve for this purpose. Stadler (1989) ex-
influence the temporal distinctiveness of the items at the tended the result and verified that the learning was implicit
time of recall. The temporal distinctiveness is higher under and not available to subjects awareness.
Bjork and Whittens conditions for the last few item pairs Cleeremans and McClelland (1991) demonstrated more
than for previous item pairs, and the same can be said of the precisely that a sequence of four items in limited-capacity
last few list items within the immediate recall conditions storage at one time may be enough to allow learning of the
that Watkins (1974) had considered. Although the long- contingencies between those items. The task was to press
term recency effect challenges a time-limited memory ex- one of six keys corresponding to stimuli at six screen loca-
planation of the recency effect, it need not challenge a tem- tions. These locations were stimulated according to a finite
porary memory capacity limit. A capacity-limited store state grammar on 85% of the trials, of which subjects were
could work in combination with the distinctiveness princi- unaware. However, on the remaining, randomly selected
ples. The recall process could proceed in phases, each of 15% of the trials, the expectations according to the gram-
which may involve the subject scanning the memory repre- mar were violated. The nature of the grammar was such
sentation, transferring several items to the capacity-limited that a prediction could be made only if one took into ac-
store, recalling those items, and then returning to the rep- count a sequence of several previous stimuli. A sequential
resentation for another limited-capacity handful of items. analysis of the reaction times showed that, after 13 experi-
It would make sense for the subject to recall the most re- mental sessions, subjects became able to use a series of
cent, most distinctive items in the first retrieval cycle so as three stimuli to predict the location of the next, fourth stim-
to avoid losing the distinctiveness advantage of those items. ulus. Even after 20 sessions, though, they remained unable
Assuming that a capacity-limited store (presumably the fo- to use a series of four stimuli to predict the location of the
cus of attention) must intervene between a memory repre- next, fifth stimulus. Thus, four seems to be the asymptotic
sentation and recall, it is consistent with the long-term re- value. (The authors presented a different theoretical ac-
cency effect. count based on the diminishing predictive value of remote
associations. However, the two accounts may not be mutu-
3.4.5. Sequential effects in implicit learning and memory. ally exclusive.)
Implicit learning is a process in which information is McKone (1995) demonstrated a capacity limit in the se-
learned without the awareness of the subject, and implicit quences that can contribute to repetition priming in lexical
memory is learned information that is revealed in an indi- decision or word naming. Series of words or nonwords were
rect test, without the subject having been questioned explic- presented and there were repetitions of items with a vari-
itly about the information. The role of a limited-capacity able number of different items intervening between the
store in implicit learning and memory exists, though its na- two instances of the repeated word (ranging from 0 to 23
ture is not yet clear (see Frensch & Miner 1994; Nissen & intervening items). The measure of priming was a decrease
Bullemer 1987; Reber & Kotovsky 1997). It is possible that in reaction time for the repeated word, suggesting that the
the role of a limited-capacity store depends on whether representation of the first instance of that word was still ac-
learning and memory require associations between items tive in memory. McKone concluded (p. 1108) that for
that go beyond a simple chain of association between adja- words, a large short-term priming component decayed
cent items. There are data supporting this conjecture and rapidly but smoothly over the first three items intervening
suggesting that implicit learning can take place if one need between the instances of the repeated word, and then
hold no more than four items in the capacity-limited store reached a stable long-term value. This appears to be evi-
at one time. dence of a series of about four consecutive items present in
Lewicki et al. (1987) examined one situation in which a limited-capacity store at any time, though a long-term
contingencies were spread across seven trials in a row, but store also contributes to priming as shown by the asymp-
the capacity-limited store need not encompass all of those totic level of residual priming at longer repetition lags.
items. Lewicki et al. presented sets of seven trials in a row An unresolved question stemming from McKone (1995)
in which the subject was to indicate which quadrant of the is why the priming declined smoothly over the last few
screen contained the target item, the digit 6 (using a key- items. When the presentation is sequential and there is no
press response). The first six trials in the set included only deliberate effort to retain any but the current item, as in this
the target, but the seventh trial also included 35 foils dis- study, it is possible that the more recent items tend to be
tributed around the screen. Moreover, unbeknownst to the more strongly represented in the limited-capacity store. It
subject, the locations of targets on Trials 1, 3, 4, and 6, taken also is possible that some of the most recent four items

sometimes are replaced in the limited-capacity store by 3.4.7. Influence of capacity on mental addition reaction
items from elsewhere in the list or from extraneous time. Logan (1988, Experiment 4) developed a task in which
thoughts. subjects had to verify equations like B 1 3 5 E (true in
this example because E is 3 letters after B in the alphabet).
3.4.6. Influence of capacity on the properties of visual The addend could be 2, 3, 4, or 5. Practice effects in this
search. Fisher (1984) observed what appears to be a limit novel task followed a power function for addends of 2
in the ability to examine items in a visual array in search of through 4. However, for an addend of 5, the fit was much
a well-learned target item. In previous work, Shiffrin and worse, and there was a discontinuity in the learning curves
Schneider (1977) and Schneider and Shiffrin (1977) distin- after about 24 presentations in which the reaction times for
guished between variably-mapped searches (in which a foil this addend dropped sharply. This discontinuity was linked
on one trial might become the target on another trial) and to a strategy shift that many subjects reported. They re-
consistently-mapped searches (in which the items that ported that problems with an addend of 5 were much
serve as targets never serve as foils). On each trial in their harder and led to a more deliberate learning strategy in
experiments, the subject knows what target or targets to which particular instances were memorized. Logan and
search for, and indicates as rapidly and accurately as possi- Klapp (1991) replicated this finding. It might be further
ble the presence or absence of the target item(s) in a visual speculated that the discontinuity could occur because the
array of items. It was proposed that variably-mapped numbers 14 can be visualized more clearly during the
searches require capacity-limited or controlled process- problem, serving as a place-holder during the adding
ing, whereas over many trials, a consistently-mapped search process. Numbers of 5 and above may be difficult because
task comes to be performed automatically, without using the visualization of items to be added is hindered by the ca-
controlled processing. Under those circumstances, it was pacity limitation.
shown that processing took place on all items in parallel, so
that the reaction time to detect a target item was nearly un- 3.4.8. Mathematical modeling parameters. Attesting to the
affected by the number of items in the array. potential importance of the pure capacity limit, various ar-
Fisher (1984) proposed that capacity limits still might ap- ticles presenting mathematical models of various complex
pear if the required rate of perceptual processing were fast cognitive processes have used the assumption that four
enough. He re-examined this assumption using a task in items can be saved in a short-term store. These include, for
which there were 20 stimulus arrays in rapid succession on example, the Kintsch and van Dijk (1978) model of text
each trial (with a pattern mask preceding the sequence of comprehension, the Raaijmakers and Shiffrin (1981) model
20 and another following it). The duration of each array var- of memory search (SAM), and the recent model of pro-
ied across trial blocks (with 10 durations between 40 and cessing capacity by Halford et al. (1998). These models pre-
200 msec per array), as did the array size (with 4 or 8 stim- sumably use a limit of four because it maximizes the ability
uli per array). The arrays were letters except for a single tar- of the model to fit real data.
get item within one array, which was always the digit 5.
The task was to indicate the spatial location of this target
item out of eight possible locations. This is a type of consis- 3.5. Empirical summary
tently-mapped task situation in which little practice is In this review, care was taken to exclude situations in which
needed to achieve an automatic search because the digit the chunking of items was unclear (yielding compound
and letter categories have been learned before the experi- STM estimates). The results of the surviving experimental
ment. The results were analyzed in light of a mathematical situations, a wide variety of situations in fact, suggest that
model, the steady-state limited-channel model, based on about four chunks can be held in a pure capacity-limited
the following defining assumptions (Fisher 1984, p. 453): STM (presumably the focus of attention). The experimen-
(1) Encoded stimuli in the visual cortex are scanned once for tal means for groups of adults generally range from about 3
placement on a comparison channel; (2) the time between ar- to 5 chunks, whereas individual subject means range more
rivals of stimuli to the comparison channels is exponentially dis- widely from 2 to 6 chunks.
tributed with rate parameter l; (3) the time to compare a stim-
ulus with a prespecified target is exponentially distributed with
rate parameter m; (4) at most k comparison channels can exe- 3.6. Testability of the theoretical analysis
cute in parallel; (5) stimuli in iconic memory are equally likely
to be replaced by the characters or masks which appear next to There are several ways in which, in future research, one
the input streams; (6) masks are not placed on the comparison could invalidate the capacity estimates of 3 to 5 chunks that
channels; and (7) the system is in a steady state. Note that it is have been derived from the many theoretical phenomena
assumed that the two dimensional coordinates of a stimulus are described above. First, one could show that performance in
retained in the visual cortex. these studies results from a grouping process in which mul-
The critical idea is that the representation of a stimulus is tiple items contribute to a chunk. In many cases the argu-
lost if that stimulus is offered to the comparison process at ment against this was limited to a theoretical rationale why
a time when there are no comparison channels free. Thus, such chunking should be absent in particular circumstances
the capacity limit in this situation is defined by the number (see sect 1.2); few studies actually have provided direct ev-
of comparison channels, k. Within the field of queuing the- idence of chunk size. Second, one could develop conditions
ory, Erlangs Loss Formula describes the problem and, us- in which more is done to limit chunking and find smaller
ing that formula, Fisher found that the data fit the formula capacity estimates. Third, one could find that there are
best with k 5 4. Assuming that the comparison channels hidden storage demands and that, when they are elimi-
are actually slots within some short-term storage mecha- nated, substantially larger capacity estimates arise (e.g.,
nism, this result serves as another indication of its limited- . 6 chunks). Fourth, in a different vein, one could find low
capacity nature (also see Schweickert et al. 1996). or zero correlations between different estimates of storage

capacity despite large individual differences in the esti- search for a particular letter in the unorganized list, one
mates. This last finding would not necessarily challenge the would search on the average through 4.5 of the items (the
notion that there are fixed capacity limits in a given domain, average of the numbers 1 through 8). If one had to search
but it would challenge the concept of a central capacity through the list organized into two chunks, one would have
mechanism (e.g., the focus of attention) that is the seat of to search on the average through 1.5 chunks to find the
the capacity limit across all domains. right chunk and then an average of 2.5 items within that
chunk to find the right item, or 4.0 accesses in all. On the
average, the hierarchical organization would be more effi-
4. Theoretical account of capacity limits:
cient. With an exhaustive search method, if one had to
Unresolved issues
search for a particular letter in the unorganized list, one
Below, I will address several fundamental theoretical ques- would have to search through eight items. For the orga-
tions about capacity limits. (1) Why does the capacity limit nized list, one would need two searches to find the right
occur? (2) What is the nature of this limit: Is there a single chunk and then four searches to find the right item within
capacity limit or are there multiple limits? (3) What are the that chunk, or six accesses in all. On the average, again, the
implications of the present arguments for alternative theo- organized list would be more efficient. In contrast, consider
retical accounts? (4) Finally, what are the boundaries of the a self-terminating search for a list of four items that could
central-capacity-limit account? An enigma in Miller (1956), be represented in an unorganized manner or as 2 chunks of
regarding absolute judgments, will be touched upon to ex- 2 items each. The unorganized list would require an aver-
amine the potential breadth of the present framework. age of 2.5 searches whereas the organized list would require
that 1.5 clusters and 1.5 items in that cluster be examined,
for a total of 3.0 searches. In this case, the unorganized list
4.1. Why the capacity limit? is more efficient on average. MacGregor calculated that or-
Future research must establish why there is a limit in ca- ganizing list items into higher-level chunks is beneficial
pacity. One possible reason is that the capacity limit has be- with an exhaustive or a self-terminating search when there
come optimized through adaptive processes in evolution. are more than 4 or 5.83 items, respectively.
Two relevant teleological arguments can be made on logi- Although these theoretical findings depend on some
cal grounds. Recently, as well, arguments have been made untested assumptions (e.g., that the difficulty of search is
concerning the physiological basis of storage capacity lim- the same at every level of a hierarchy), they do provide use-
its. Any such account must consider details of the data in- ful insight. The empirically observed capacity limit of about
cluding the individual variability in the capacity limit esti- four chunks corresponds to what has been predicted for
mates that have been observed as discussed in section 3.1.3. how many items can be advantageously processed in an un-
These issues will be addressed in turn. grouped manner when the search is exhaustive (MacGre-
gor 1987). These theoretical and empirical limits may cor-
4.1.1. Teleological accounts. Several investigators have respond because very rapid searches of unorganized lists
provided mathematical arguments relevant to what the are, in fact, exhaustive (Sternberg 1966). However, slower,
most efficient size of working memory would be. Dirlam self-terminating searches along with more elaborate men-
(1972) asked if there is any one chunk size that is more ef- tal organization of the material also may be possible, and
ficient than any other if it is to be the basis of a memory probably are advantageous if there is time to accomplish
search. He assumed that STM is a multi-level, hierarchi- this mental organization. That possibility can help to ex-
cally structured system and that the search process is ran- plain why the empirically observed limit of about four
dom. The nodes at a particular level are searched only un- chunks is close to the optimal chunk size when multiple lev-
til the correct one is found, at which time the search is els of organization are permitted in a self-terminating
confined to subnodes within that node. This process was as- search (Dirlam 1972).
sumed to continue until, at the lowest level of the hierar- Another teleological analysis can be formulated on the
chy, the searched-for item is identified. In other words, the basis of Kareev (1995). He suggested that a limited work-
search was said to be self-terminating at each level of the ing memory is better than an unlimited one for detecting
hierarchy. Dirlam then asked what rule of chunking would imperfect correlations between features in the environ-
minimize the expected number of total node and item ac- ment. To take a hypothetical example, there could be a pop-
cesses regardless of the number of items in the list, and cal- ulation with a 70% correlation between the height of an in-
culated that the minimum would occur with an average dividual and the pitch of his or her voice. In a statistical
chunk size of 3.59 items at each level of the hierarchy, in procedure, when one uses a limited sample size to estimate
close agreement with the capacity of short-term memory the correlation (e.g., an observation of 4 8 individuals), the
that has been observed empirically in many situations (see modal value of the observed correlation is larger than the
above). population value. The smaller the sample size, the higher
MacGregor (1987) asked a slightly different question: the modal value. Thus, a smaller sample size would increase
What is the maximal number of items for which a one-level the chances that a moderate correlation would be noticed
system is more efficient than a two-level system? The con- at all. In human information processing, the limit in the
sequences of both self-terminating search and exhaustive sample size could be caused by the capacity limit of the ob-
search assumptions were examined. A concrete example servers short-term memory; more samples may have been
would help to explain how. Suppose that one received a list observed but the observer bases his or her perceived esti-
that included eight items. Further suppose that one had the mate of the correlation on only the number of examples that
option of representing this list either in an unorganized fit into the focus of attention at one time. Kareev et al.
manner or as two higher-level chunks, each containing four (1997) showed that, in fact, low-working-memory subjects
items. With a self-terminating search method, if one had to were more likely to notice a population correlation of .2.6.

In this regard, it bears mention that in the statistical sam- stimulus was presented when an object (a silhouetted hu-
pling procedure, the modal value of the sample correlations man profile) was perceived, but not when a random field
for sample sizes of 6 and 8 were shown to be only moder- (actually an upside down profile not detected as such) was
ately greater than the true population value (which was set perceived. The scalp locations appeared to implicate the
at .6 or .7); but for a sample size of 4, the modal value of the parietal lobes, which Cowan (1995) also proposed to be ar-
sample correlations was almost 1.0. Here, then, is another eas involved in the integration of features to form objects.
reason to believe that a basic capacity limit of four could be Miltner et al. (1999) further showed that the binding can
advantageous. It could take a moderate correlation in the take place not only between perceptual features, but also
real world and turn it into a perceived strong correlation. At between a feature and an activated mental concept. Specif-
least, this could be advantageous to the extent that deci- ically, cyclic activity in the gamma (2070 Hz) band was
siveness in decision-making and definiteness in the percep- synchronized between several areas of the brain in the time
tion of stimulus relationships are advantageous. For exam- period after the presentation of a conditioned stimulus
ple, it makes sense to walk away from someone displaying (CS1), a color illuminating the room, but before the pre-
traits that are moderately correlated with injurious behav- sentation of the unconditioned stimulus (UCS), an electric
ior, and it makes sense to perceive that people usually say shock that, as the subjects had learned, followed the condi-
please when they are asking for a favor. tioned stimulus. No such synchronization occurred after a
There is a strong similarity between the theoretical analy- different color (CS-) that did not lead to electric shock.
sis of Kareev and earlier proposals that a large short-term If objects and meaningful events can be carried in the syn-
memory capacity can be a liability rather than a strength in chronized activity of gamma wave activity in the brain, then
the early stages of language learning. Newport (1990) dis- the question for STM capacity becomes, How many objects
cussed a less is more hypothesis to explain why language or events can be represented simultaneously in the brain?
learners who are developmentally immature at the time of Investigators have discussed that. Lisman and Idiart (1995)
initial learning have an advantage over more mature learn- suggested that each memory is stored in a different high-
ers for some language constructs. An alternative to the na- frequency (40 Hertz) subcycle of a low-frequency oscilla-
tivist theory of language learning, this theory states that im- tion. Memory patterns repeat on each low-frequency (5 to
mature language learners grasp only small fragments of 12 Hertz) oscillation, a repetition that relies on activity de-
language at a time, which helps them to break up a complex pendent changes in membrane excitability rather than re-
language structure into smaller parts. Consistent with this verberatory circuits. In other words, the number of subcy-
proposal, Elman (1993) found that a computer implemen- cles that fit into a low-frequency cycle would define the
tation of a parallel distributed processing model of cogni- number of items that could be held in a capacity-limited
tion learned complex language structure more easily if the STM. This suggestion was intended by Lisman and Idiart to
short-term memory capacity of the model started out small motivate the existence of a memory span of about seven
and increased later in the learning process, rather than tak- items (e.g., [40 subcycles/sec] / [5.7 cycles/sec] 5 7 subcy-
ing on its mature value at the beginning of learning. cles/cycle). However, it could just as well be used to moti-
Below, neurophysiological accounts of capacity limits will vate a basic capacity of about 4 items (e.g., [40 subcycles/
be reviewed. The teleological arguments still will be im- sec]/[10 cycles/sec] 5 4 subcycles/cycle). This proposal also
portant to the extent that they can be seen as being consis- was intended to account for the speed of retrieval of infor-
tent with the physiological mechanisms underlying capac- mation stored in the capacity-limited STM but, again, just as
ity limits (or, better yet, motivating them). well fits the 4-item limit. If 40 subcycles occur per second
then each subcycle takes 25 msec, a fair estimate of the time
4.1.2. Neurophysiological accounts. In recent years, a it takes to search one item in STM (Sternberg 1966). Luck
number of investigators have suggested a basis of capacity and Vogel (1998) made a proposal similar to Lisman and
limits that can be traced back to information about how a Idiart but made it explicit that the representation of each
single object is represented in the brain. In a theoretical ar- item in STM would involve the synchronization of neural fir-
ticle on visual shape recognition, Milner (1974, p. 532) sug- ing representing the features of the item. The STM capac-
gested that cells fired by the same figure fire together but ity limit would occur because two sets of feature detectors
not in synchrony with cells fired by other figures . . . Thus, that fire simultaneously produce a spurious synchronization
features from a number of figures could be detected and corrupting memory by seeming to come from one object.
transmitted through the network with little mutual inter- Other theorists (Hummel & Holyoak 1997; Shastri & Aj-
ference, by a sort of time-sharing arrangement. In support janagadde 1993) have applied this neural synchronization
of this hypothesis, Gray et al. (1989), in an experiment on principle in a way that is more abstract. It can serve as an
cats, found that two columns of cortical cells that repre- alternative compatible with Halford et al.s (1998) basic no-
sented different portions of the visual field were active in a tion of a limit on the complexity of relations between con-
correlated manner only if they were stimulated by different cepts, though Halford et al. instead worked with a more
portions of the same object, and not if they were stimulated symbolically based model in which the amount of infor-
by different objects. This led to the hypothesis that the syn- mation that can be represented by a single vector is not sig-
chronization of activity for various features represents the nificantly limited, but the number of vectors that can be
binding of those features to form an object in perception or bound in one representation of a relation is limited
STM. More recently, these findings have been extended to (p. 821). Shastri and Ajjanagadde (1993) formulated a phys-
humans. Tiitinen et al. (1993) found that the 40-Hz oscilla- iological theory of working memory very similar to Lisman
tory cycle, upon which these synchronizations are thought and Idiart (1995), except that the theory was meant to ex-
to ride, is enhanced by attention in humans. Rodriguez et plain a limited-capacity dynamic working memory that
al. (1999) reported electrical synchronizations between cer- temporarily holds information during an episode of reflex-
tain widely separated scalp locations 180 360 msec after a ive reasoning (p. 442), meaning reasoning that can be car-

ried out rapidly, spontaneously, and without conscious ef- slow oscillation is close to about 10 Hz, so that four items
fort (p. 418). The information was said to be held as con- would fit in each of the slower cycles. As suggested by Shas-
cepts or predicates that were in the form of complex tri and Ajjanagadde (1993) and others, the cyclic search
chunks; thus, it was cautioned, note that the activation of process could be employed recursively. For example, at one
an entity together with all its active superconcepts counts point in a probed recognition process there could be up to
as only one entity (p. 443). It was remarked that the bound four chunks in the capacity-limited store. Once the correct
on the number of entities in working memory, derived from chunk is identified, the contents of STM would be replaced
facts of neural oscillation, falls in the 7 6 2 range; but the by the items contained within that chunk, now unpacked,
argument was not precise enough to distinguish that from so that the contents of the chunk can be scanned in detail.
the lower estimate offered in the present paper. Hummel In present theoretical terms, the focus of attention need not
and Holyoak (1997) brought up similar concepts in their focus on multiple levels of representation at the same time.
theory of thinking with analogies. They defined dynamic
binding (a term that Shastri & Ajjanagadde also relied 4.1.4. What is the basis of individual differences? We will
upon to describe how entities came about) as a situation in not have a good understanding of capacity limits until we
which units representing case roles are temporarily bound are able to understand the basis of the marked develop-
to units representing the fillers of those roles (p. 433). mental and individual differences in measured capacity that
They estimated the limit of dynamic binding links as be- were observed by Cowan et al. (1999) and comparable in-
tween four and six (p. 434). In both the approaches of dividual differences observed in other procedures (Hen-
Shastri and Ajjanagadde (1993) and Hummel and Holyoak derson 1972; Luck & Vogel, personal communication, Jan-
(1997), these small limits were supplemented with data uary 18, 1999). One possible basis would be individual
structures in long term memory or static bindings that ap- differences in the ratio of slow to fast oscillatory rhythms.
pear to operate in the same manner as the long-term work- Miltner et al. (1999) found most rapid oscillatory activity at
ing memory of Ericsson and Kintsch (1995), presumably 3743 Hz, but some residual activity at 3037 and 4348
providing the active superconcepts that Shastri and Aj- Hz. One can combine a 12-Hz slow cycle with a 30-Hz rapid
janagadde mentioned. cycle to predict the low end of the range of memory capac-
One problem for the interpretation of synchronous os- ities (12/30 5 2.5 items), or one can combine an 8-Hz slow
cillations of nervous tissue is that they can be observed even cycle with a 48-Hz fast cycle to predict the high end of the
in lower animals in situations that appear to have little to do range (8/48 5 6 items). According to these figures, how-
with the possibility of conscious awareness of particular stim- ever, one would not expect the slow cycle to go below 8 Hz,
uli (e.g., Braun et al. 1994; Kirschfeld 1992). This, in itself, given the capacity limits observed empirically. Here, then,
need not invalidate the role of oscillations in binding to- is a physiological prediction based on a combination of ex-
gether the features of an object or the objects in a capacity- isting physiological and behavioral results. An important
limited store in humans. It could be the case that mecha- next step may be the acquisition of data that can help to
nisms already present in lower animals form the basis of evaluate the psychological plausibility of the theoretical
more advanced skills in more advanced animals, just as the constructs surrounding this type of theory. As one promis-
voice apparatus is necessary for speech but is present even ing example, the finding of Tiitinen et al. (1993) that the 40-
in non-speaking species. Thus, von der Malsburg (1995, p. Hz neural cycle is enhanced by attention is consistent with
524) noted that As to the binding mechanism based on the present suggestion that the fundamental storage capac-
temporal signal correlations, its great advantage [is] being ity limit of about four items is based on the 40-Hz cycle and
undemanding in terms of structural requirements and con- is in essence a limit in the capacity of the focus of attention.
sequently ubiquitously available and extremely flexible. It is easy to see how research on this topic also could clar-
ify the basis of individual differences in capacity. Specifi-
4.1.3. Reconciliation of teleological and neurophysiolog- cally, one could determine if individual differences in oscil-
ical accounts. One concern here is whether the teleologi- latory rates mirror behavioral differences in the limited
cal and physiological accounts of capacity limits are consis- storage capacity.
tent or inconsistent with one another. The process of It remains to be explained why attended speech shows
scanning through the items in STM has been employed the- such an intriguing, simple relationship to unattended
oretically by both the teleological and the physiological the- speech (Fig. 4) in which attended speech is increased above
orists. For example, the teleological argument that Mac- the unattended speech limit by a variable amount. This fig-
Gregor (1987) built using an exhaustive scan resulted in the ure makes it apparent that individuals use the same pro-
conclusion that the scan would be most efficient if the num- cesses in both conditions, plus supplementary processes for
ber of items per group were four. This conclusion was based attended speech. This difference might be accounted for
on the assumption that the amount of time it takes to access most simply by the process of chunking (formation of inter-
a group to determine whether a particular item is present item associations) during attended list presentations.
within it is equal to the amount of time it then takes to ac- It is important not to become too reductionistic in the in-
cess each item within the appropriate group once that terpretation of biological effects. It is possible that stimulus
group is selected, so as finally to identify the probed item. factors and/or behavioral states modulate biological cycle
This concept can be mapped directly onto the concept of frequencies under some circumstances. Some studies with
the set of items (or chunks) in capacity-limited STM being an automatized response or a rapid response have resulted
represented by a single cycle of a low-frequency oscillation in smaller individual differences. The highly trained sub-
(5 to 12 Hz) with each item mapped onto a different cycle jects of Sperling (1960) appeared to produce capacity
of a 40-Hz oscillation, riding on top of the 5 to 12 Hz oscil- (whole report) estimates deviating from the population
lation. These figures are in line with the teleological data mean by no more than about 0.5 items, although there were
and memory capacity data reviewed above if the rate for the few subjects. In an enumeration task in which a reaction

time measure defined the subitizing range, Chi and Klahr Henderson 1972; Klapp & Netick 1988; Luck & Vogel
(1975) found no difference between 5- and 6-year-olds ver- 1997; Martin 1980; Penney 1980; Reisberg et al. 1984;
sus adults in the subitizing range. Perhaps there is an in- Sanders & Schroots 1969; Shah & Miyake 1996).
trinsic, baseline capacity of the focus of attention that shows Theoretically, it should be possible to overcome method-
few differences between individuals, and perhaps under ological problems in order to determine if there are true
some circumstances but not others, the level and direction multiple capacity limits. One could make it impossible for
of effort at the time of recall modulate that capacity. Fur- the subject to rehearse items during presentation of the ma-
ther study of individual differences in memory capacity is terials by using complex arrays of two types concurrently;
thus likely to be important theoretically. perhaps concurrent visual and auditory arrays. It would also
be necessary to make sure that the focus of attention could
not be used recursively, shifting from one type of activated
4.2. Central capacity or separate capacities?
material to the next for recall. If the activated representa-
In most of the research that I have discussed, the capacity tions were sensory in nature, this recursive recall might be
limit is examined with a coherent field of stimulation. I have prevented simply by backward-masking one or both of the
not directly tackled the question of whether there is one types of materials. These requirements do not seem to have
central capacity limit or whether there are separate limits been met in any extant study. Martin (1980) did use simul-
for domains of cognition (e.g., separate capacities for the vi- taneous left- and right-sided visual and auditory lists (4
sual versus auditory modalities; for verbal versus spatial channels at once, only 2 of them meaningful at once, with
representational codes; or, perhaps, for two collections of sequences of 4 stimuli presented at a fast, 2/sec rate on each
items distinguished by various other physical or semantic of the 4 channels). She found that memory for words pre-
features). According to the models of Cowan (1988; 1995) sented concurrently to the left and right fields in the same
and Engle et al. (1999), the capacity limit would be a cen- modality was, on the average, 51.6% correct, whereas mem-
tral one (the focus of attention). Some fine points must be ory for pairs containing one printed and one spoken word
kept in mind on what should count as evidence for or was 76.9% correct. However, there was nothing to prevent
against a central limit. the shifting of attention from visual to auditory sensory
Ideally, evidence for or against a central capacity limit memory in turn.
could be obtained by looking at the number of items re- Another methodological possibility is to document the
called in two tasks, A and B, and then determining whether shifting of attention rather than preventing it. This can be
the total number of items recalled on a trial can be in- accomplished with reaction time measures. One enumera-
creased by presenting A and B together and adding the tion study is relevant. Atkinson et al. (1976b) presented two
number of items recalled in the two tasks. For example, sets of dots separated by their organization into lines at dif-
suppose that one can recall three items in Task A and four ferent orientations, by two different colors, or by their or-
items in Task B, and that one can recall six items all together ganization into separate groups. Separation by spatial ori-
in a combined, A 1 B task. Although performance on the entation or grouping was capable of eliminating errors
component tasks is diminished when the tasks are carried when there was a total of five to eight dots. Color separa-
out together, the total number of items recalled is greater tion reduced, but did not eliminate, errors. However, the
than for either task presented alone. This savings would grouping did not reduce the reaction times in any of these
serve as initial evidence for the existence of separate stor- studies. It seems likely that some sort of apprehension
age mechanisms (with or without an additional, central process took place separately for each group of four or
storage mechanism). Further, if there were no diminution fewer dots and that the numbers of dots in each group were
of performance in either task when they were combined, then added together. Inasmuch as the reaction times were
that would serve as evidence against the central storage not slower when the items were grouped, one reasonable
mechanism or capacity limit. interpretation is that subitizing in groups and then adding
This type of reasoning can be used only with impor- the groups is the normal enumeration process for fields of
tant limitations, however. As discussed above, several dif- five or more dots, even when there are no physical cues for
ferent mechanisms contribute to recall, including not only the groups. The addition of physical cues simply makes the
the capacity-limited focus of attention, but also the time- subitizing process more accurate (though not faster). This
or interference-limited sources of activation of long-term study provides some support for Mandlers (1985) sugges-
memory (sensory stores, phonological and spatial stores, tion that the capacity limit is for sets of items that can be
etc.). If the focus of attention could shift from examining one combined into a coherent scheme. By dividing the sensory
source of activation to examining another dissimilar source, field into two coherent, separable schemes, the effective
it would be possible to recall items from Task A and then limit can be increased; but different schemes or groups can
shift attention to activated memory representations of the become the limit of attention only one at a time, explaining
items in Task B, bringing them into the focus of attention why perceptual grouping cues increase accuracy without al-
for recall in turn. If all of the information need not be en- tering the reaction times.
tered into the focus of attention at one time, performance Physiological studies also may help if they can show a rec-
in the combined task would overestimate central storage iprocity between tasks that do not appear to share specific
capacity. This possibility contaminates many types of evi- processing modes. One study using event-related potentials
dence that initially look as if they could provide support (ERPs) by Sirevaag et al. (1989) is relevant. It involved two
for multiple capacity-limited stores. These include various tasks with very little in common, both of which were ef-
studies showing that one can recall more in two tasks with fortful. In one task, the subject controlled a cursor using a
different types of materials combined than in a single task, joystick, in an attempt to track the movement of a moving
especially if the modalities or types of representations are target. The movement could be in one or two dimensions,
very different (Baddeley 1986; Frick 1984; Greene 1989; always in discrete jumps, and the cursor could be controlled

by either the velocity or the acceleration of the joystick, re- model of Baddeley (1986). The research reviewed in the
sulting in four levels of task difficulty. In the second task, present article leaves open the question of whether time
administered concurrently, the subject heard a series of limitations exist (as explained in sect. 1). Whereas some
high and low tones and was to count the number of occur- have assumed that time limits can take the place of capac-
rences of one of the tones. The P300 component of ERP ity limits, the evidence described in this article, however,
responses to both tasks was measured. This component is cannot be explained in this manner. In Baddeleys theory,
very attention-dependent. The finding was that, across con- memory span was said to be limited to the number of items
ditions, the P300 to the tracking targets and the P300 to the that could be rehearsed in a repeating loop before their rep-
tones exhibited a reciprocity. The larger the P300 was to the resentations decay from the storage buffer in which they
tracking targets, the smaller it was to the tones, and vice are held (in about 2 sec in the absence of rehearsal). If re-
versa. The sum of the P300 amplitudes was practically con- hearsal is always articulatory in nature, though, this notion
stant across conditions. The simplest interpretation of these is inconsistent with findings that the memory span for id-
results is that there is a fixed capacity that can be divided ioms would imply a much longer rehearsal time than the
among the two tasks in different proportions, and that the memory span for individual characters (e.g., see Glanzer &
relative P300 amplitudes reflect these proportions. Razel 1974). Something other than just the memorys dura-
In sum, the existing literature can be accounted for with tion and the rate of articulatory rehearsal must limit recall.
the hypothesis that there is a single capacity-limited store The time-based account might be revived if a different
that can be identified with the focus of attention. This store means of rehearsal could be employed for idioms than for
is supplemented with other storage mechanisms that are words. For example, subjects might be able to scan seman-
not capacity limited although they are limited by the pas- tic nodes, each representing an idiom, and quickly reacti-
sage of time and/or the presentation of similar interfering vate them in that way without articulatory rehearsal of the
material. The focus of attention can shift from one type of idioms. According to a modified version of Baddeleys ac-
activated memory to another and will recoup considerable count, this scanning would have to be completed in about
information from each type if the materials represented are 2 sec to prevent decay of the original memory traces. How-
dissimilar. ever, even that modified time-based theory seems inade-
quate to account for situations in which the material to be
recalled is presented in an array so quickly that rehearsal of
4.3. Implications for alternative accounts of information any kind can contribute little to performance (e.g., Luck &
processing Vogel 1997). Also, any strictly time-based account has diffi-
In light of the information and ideas that have been pre- culty explaining why there is an asymptotic level of recall in
sented, it is important to reconsider alternative accounts of partial report approximating four items with both auditory
information processing and the question of their continued and visual presentation of characters, even though it takes
viability and plausibility. much longer to reach that asymptote in audition (at least 4
sec: Darwin et al. 1972) than in vision (at most 1 sec: Sper-
4.3.1. The magical number seven, plus or minus two. Al- ling 1960). The only way to preserve a time-based account
though Miller (1956) offered his magical number only as a would be to assume that the rate of extraction of informa-
rhetorical device, the number did serve to characterize per- tion from sensory storage in the two modalities is a limiting
formance in many tasks. It has been taken more literally as factor and is, for some mysterious reason, inversely pro-
a memory limit by many researchers (e.g., Lisman & Idiart portional to the duration of sensory storage, resulting in an
1995). The present stance is that the number seven esti- asymptotic limit that does not depend on the duration of
mates a commonly obtained, compound capacity limit, storage. It seems far simpler to assume a capacity limit.
rather than a pure capacity limit in which chunking has
been eliminated. It occurs in circumstances in which the 4.3.3. The unitary storage account. Some theorists (e.g.,
stimuli are individually attended at the time of encoding Crowder 1993) have assumed that there is no special short-
and steps have not been taken to eliminate chunking. What term memory mechanism and that all memory may be ex-
is needed, however, is an explanation of why this particular plained according to a common set of rules. In one sense
compound limit crops up fairly often when rehearsal is not the present analysis is compatible with this view, in that the
prevented. One possibility is that this number reflects a cer- capacity limit applies not only to the recall of recently pre-
tain reasonable degree of chunking. Most adults might be sented stimuli, but also to the recall of information from
able to learn at most three chunks of information rapidly, long-term memory (see sect. 3.4.2). However, any success-
each with perhaps three units, leading to a span of nine. The ful account must distinguish between the vast information
slightly lower estimates that are often obtained could result potentially obtainable from an individual, on one hand, and
from the inability to learn the chunks quickly enough. How- the small amount of information that can be obtained from
ever, these speculations are intended only to provoke fur- that individual, or registered with the individual, in a short
ther research into the basis of commonly obtained com- segment of time; the capacity limit. The focus of attention,
pound capacity limits. What is essential to point out in the which serves as the proposed basis of the capacity limit in
present account is that these compound limits are too high the present approach, has not played a major role in unitary
to describe performance in the situations in which it can be accounts that have been put forward to date, though it
assumed that the combination of items into higher-order could be added without contradiction.
chunks was severely limited or prevented. Given a unitary memory view expanded to consider the
focus of attention, one could account for the 4-chunk limit
4.3.2. The time-limitation account. The view that working on the grounds that every chunk added to the focus dimin-
memory is limited by the duration of unrehearsed informa- ishes the distinctiveness of all the chunks. Such a mecha-
tion in various short-term buffers is exemplified by the nism of indistinctiveness would be analogous to the one that

has been used previously to account for the recency effect 4.3.6. The storage versus processing capacities account.
in serial recall (e.g., Bjork & Whitten 1974); except that the Daneman and Carpenter (1980), like many other investi-
dimension of similarity between chunks would be their con- gators, have noted that a working memory storage load does
current presence in the focus of attention, not their adja- not interfere with processing nearly as much as would be
cent serial locations within a list. One article written from a expected if storage and processing relied upon a common
unitary memory view (Brown et al. 2000) does attempt to workspace. Halford et al. (1998) noted the storage limit of
account for the number of chunks available in one situation. about four items but also proposed, parallel to that limit but
Specifically, their account, based on oscillatory rhythms separate from it, a processing limit in which the complexity
that become associated with items and contexts, correctly of relations between items being processed is limited to
predicted that the serial recall of a nine-item list with overt four dimensions in adults (and to fewer dimensions in chil-
rehearsal is optimal when the list is rehearsed in groups of dren). Thus, within processing, complexity is defined as
three items. The explanation offered was that This repre- the number of related dimensions or sources of variation
sents the point at which the optimal balance between (p. 803). For example, transitive inference is said to be a
across-group errors and within-group errors is reached in ternary relation because it can be reduced to such terms:
the model. The account of the 4-chunk limit offered ear- the premises Tom is smarter than John, John is smarter
lier in this target article on the basis of neural oscillatory than Stan can be integrated into the ternary relational in-
rhythms (sect. 4.1.2) is similar (albeit on a neural level of stance monotonically-smarter (Tom, John, Stan) (p. 821),
analysis). It states that, with the neural representation of too an argument with three fillers. The parallel between pro-
many chunks simultaneously, the representations begin to cessing and storage was said to be that both attributes on
become confusable (cf. Luck & Vogel 1998). The critical dimensions [in processing] and chunks [in storage] are in-
difference between explanations is that the neural account dependent units of arbitrary size (p. 803). However, the
offered by Luck and Vogel and the present article refers to model did not explain why there was the coincidental sim-
particular frequencies of neural oscillation, whereas Brown ilarity in the processing and storage limits, to about four
et al. allowed various oscillators and did not make predic- units each.
tions constrained to particular frequencies of oscillation. According to the present view, both processing and stor-
age would be assumed to rely on a common capacity limit.
4.3.4. The scheduling account. It has been proposed that The reason is that, ultimately, what we take to be stored
supposed capacity limits might be attributable to limits in chunks in short-term memory (and what I have, for sim-
the rate at which subjects can produce responses in a multi- plicity, described as such up to this point) actually are rela-
task situation without risking making responses in the in- tions between chunks. It is not chunks per se that have to
correct order (Meyer & Kieras 1997). That theory appears be held in short-term memory (as they in fact are part of
more applicable to some situations than to others. In situa- long-term memory), but rather chunks in relation to some
tions in which the limit occurs during reception of materi- concept. For example, in-present-array (x, q, r, b) could
als and fast responding is not required (e.g., Luck & Vogel describe the quaternary relation leading to a whole report
1997), the theory seems inappropriate. That seems to be response in Sperlings (1960) procedure. Monotonically-
the case with most of the types of phenomena examined in later (37, x, 2, 48) could describe a quaternary relation
the present article. It is unclear how a scheduling account leading to partially correct serial recall of an attended list of
could explain these phenomena without invoking a capac- digits for which 3 7 is a memorized initial chunk; x repre-
ity notion. sents a placemarker for a digit that cannot be recalled; 2
represents an unchunked digit; and 4 8 represents another
4.3.5. The multiple-capacity account. Some theorists have memorized chunk.
suggested that there is not a single capacity limit, but rather If this analysis is correct, there is no reason to expect a
limits in separate capacities (e.g., visual and auditory or spa- separation between processing and storage. The reason
tial and verbal; see Wickens 1984). I would suggest that, al- why a storage load does not much interfere with processing
though there may well be various types of distinct processes is that the storage load and the process do not have to be
and storage facilities in the human brain, there is no evi- expanded in the focus of attention at the same time. Al-
dence that they are limited by capacity per se (as opposed though both are activated at the same time, there is no ca-
to other limitations such as those imposed by decay and in- pacity limit on this activation, only with its use (cf. Schnei-
terference). Sections 1 and 3 of the present article should der & Detweiler 1987). The subject might only hold in the
illustrate that strict conditions must apply in order for focus of attention a pointer to the activated, stored infor-
chunk capacity limits to be clearly observed at all, free of mation while carrying out the processing, and then the sub-
other factors. Moreover, the finding of Sirevaag et al. (1989) ject could shift the focus to the stored information when
of a tradeoff between tasks in the the P300 response mag- necessary to recall the memory load.
nitude in event-related potentials (discussed in sect. 4.2)
seems to indicate that very disparate types of processes still 4.3.7. The task-specific capacities account. A skeptic
tap a common resource. Even the left and right hemi- might simply assume that although there are capacity lim-
spheres do not appear to operate independently. Holtzman its, they vary from situation to situation for reasons that we
and Gazzaniga (1992) found that split brain patients are im- cannot yet understand. This type of view probably cannot
peded in responses made with one hemisphere when a con- be answered through reasoned discourse as it depends on
current load is imposed on the other hemisphere, despite a different judgment of the presented evidence. Further as-
the breakdown in informational transmission between the sessment of the view that there is a fixed underlying capac-
hemispheres through the corpus collosum. There thus ap- ity could be strengthened by subsequent research in which
pears to be some central resource that is used in disparate new conditions are tested and found to conform to or vio-
tasks, and by both hemispheres. late the capacity limit. Numerous examples of novel condi-

Commentary/Cowan: The magical number 4
tions leading to the predicted limit can be given, but two of line rehearsal, chunking, and memorization, for which the
them are as follows. First, in the research by Cowan et al. exact number of chunks in memory cannot be ascertained.
(1999), a capacity limit for ignored speech was expected to The fundamental capacity limit appears to coincide with
be similar to those that have been obtained for attended vi- conditions in which the chunks are held in the focus of at-
sual arrays (Luck & Vogel 1997; Sperling 1960) on the tention at one time; so it is the focus of attention that ap-
grounds that the task demands were logically analogous, pears to be capacity-limited.
even though the materials were very different. That expec- When the material to be remembered is diverse (e.g.,
tation was met. Second, it was expected that one could ob- some items spoken and some printed; some words and
serve the capacity limit by limiting rehearsal for spoken some tones; or some verbal and some nonverbal items), the
lists, and that expectation provided a very similar limit in scene is not coherent and multiple retrievals result in con-
numerous published experiments (as shown in Table 2). A siderably better recall. This all suggests that the focus of at-
third example has yet to be tested. Specifically, it was pre- tention, as a capacity-limited storage mechanism, can shift
dicted (in sect. 1.2.1) that the capacity limit could be ob- from one type of material to another or from one level of or-
served in a modified n-back task in which subjects must ganization to another, and that the individual is only aware
only indicate, as rapidly as possible, if a particular item has of the handful of separate units of a related type within a
been included in the stimulus set previously, and in which scene at any one moment (Cowan 1995; Mandler 1985).
some items would be repeated in the set but other, novel
items also would be introduced. ACKNOWLEDGMENTS
This project was supported by NICHD Grant R01 21338. I thank
Monica Fabiani, Gabriele Gratton, and Michael Stadler for help-
4.4. Boundaries of the central-capacity-limit account ful comments. Address correspondence to Nelson Cowan, De-
partment of Psychology, University of Missouri, 210 McAlester
The boundaries of the present type of analysis have yet to
Hall, Columbia, MO 65211, USA. Electronic mail: cowanN@
be examined. For example, Miller (1956) indicated that ab- missouri.edu.
solute judgments in perception are limited in a way that is
not clear; apparently not in chunks as for other types of phe-
nomena. For unidimensional stimuli the limit appears to be
up to about seven categories that can be used consistently,
but the limit in the number of total categories is consider-
ably higher for multidimensional stimuli (e.g., judgments of
tones differing in both intensity and pitch). One possibility Open Peer Commentary
is that the subject need only retain, in short-term memory,
pointers to the dimensions while accessing category divi-
sions one dimension at a time. Because faculties that are not Commentary submitted by the qualified professional readership of this
specifically capacity-limited, such as sensory memory, can journal will be considered for publication in a later issue as Continuing
be used for supplementary storage, the focus of attention is Commentary on this article. Integrative overviews and syntheses are es-
free to shift to allow the sequential use of the capacity limit pecially encouraged.
to judge the stimulus on different dimensions, one at a time.
This analysis might be tested with absolute judgments for
backward-masked stimuli, as backward masking would pre-
vent sensory storage from holding information while the fo- The dangers of taking capacity limits
cus of attention is shifted from one dimension to another. too literally
Thus, as in this example, the capacity concept potentially S. E. Avons, Geoff Ward, and Riccardo Russo
might have a broad scope of application indeed. Department of Psychology, University of Essex, Colchester CO7 6SJ, United
Kingdom. {savons; gdward; rrusso}@essex.ac.uk
www.essex.ac.uk/psychology/people/{sa; gw; rr}.html
5. Conclusion
Abstract: The empirical data do not unequivocally support a consistent
In this target article I have stressed several points. The first fixed capacity of four chunks. We propose an alternative account whereby
is the remarkable degree of similarity in the capacity limit capacity is limited by the precision of specifying the temporal and spatial
context in which items appear, that similar psychophysical constraints limit
in working memory observed with a wide range of proce- number estimation, and that short term memory (STM) is continuous with
dures. A restricted set of conditions is necessary to observe long term memory (LTM).
this limit. It can be observed only with procedures that al-
low assumptions about what the independent chunks are, In his target article, Cowan makes a brave attempt to unify atten-
and that limit the recursive use of the limited-capacity store tion and short term memory (STM) by stressing the similarity of
(in which it is applied first to one kind of activated repre- capacity estimates across various paradigms. His evidence is
sentation and then to another type). The preponderance of drawn from four broadly defined sources: (1) the number of cor-
evidence from procedures fitting these conditions strongly rect reports in whole report paradigms; (2) the size of memory
suggests a mean memory capacity in adults of three to five span when rehearsal and recoding are discouraged; (3) perfor-
mance discontinuities in various attentional and STM tasks; (4) a
chunks, whereas individual scores appear to range more miscellany of indirect measures including recency in free recall,
widely from about two up to about six chunks. The evidence implicit memory, and visual search. The capacity estimates cited
for this pure capacity limit is considerably more extensive by Cowan are less consistent than he claims. If four of the most
than that for the somewhat higher limit of 7 6 2 stimuli; recently presented items exist within the attentional focus, then
that higher limit is valid nevertheless as a commonly ob- recall of about four of the most recent items would be expected.
served, compound STM limit for materials that allow on- This is true for enumeration, whole report, and alphanumeric

span tasks under certain conditions. However, performance fails and LTM (one consequence of Cowans theory) is reduced, al-
to reach this level in other studies, such as recall of familiar visual though the strategies and processes useful for memory may in-
stimuli (Zhang & Simon 1985), multi-object tracking (Yantis deed change as a function of elapsed time, and the labels may be
1992), free recall of the most recently rehearsed items (Tan & convenient and pragmatically useful.
Ward, in press), and, as Cowan acknowledges, in repetition prim-
ing (McKone 1995).
Our main objection is more fundamental and takes the follow-
ing form: (a) these data indicate memory for events rather than A biocognitive approach to the conscious
items, (b) the notion of a fixed capacity limit, restricted to three or
four memory elements is misleading, and (c) that these consider-
core of immediate memory
ations, taken together, question the qualitative distinction be- Bernard J. Baars
tween STM and LTM that Cowan advocates.
The Wright Institute, Berkeley, CA 94704
Memory for events, not items (or chunks). Cowan propose that
baars@cogsci.berkeley.edu
the capacity of STM is not 7 chunks, as proposed by Miller (1956),
but about 4 chunks. The earlier estimate was inflated because it
was based on rehearsal-dependent immediate serial recall, and Abstract: The limited capacity of immediate memory rides on the even
more limited capacity of consciousness, which reflects the dynamic activ-
also because the span measure estimated the 50% probability of ity of the thalamocortical core of the brain. Recent views of the conscious
recall. In fact, many STM tasks require recall not only of recent narrow-capacity component of the brain are explored with reference to
items, but also of their temporal or spatial context. This is true for global workspace theory (Baars 1988; 1993; 1998). The radical limits of im-
studies of whole report, serial recall, and serial-order reconstruc- mediate memory must be explained in terms of biocognitive brain archi-
tion (sects. 3.1 and 3.2 of the target article), and applies also to the tecture.
Steinberg paradigm (sect. 3.3.4). The apparent limit of four items
could arise in two ways. It may be the case that only four items I am pleased that Cowan finds my global workspace theory useful
may be represented in immediate memory (i.e., within the focus in understanding the limited item capacity of immediate memory
of attention), or that contextual information can only be preserved (Baars 1988; 1997; Baars & Newman, in press). Without propos-
for about four items. Since order and position errors are common ing a specific solution to the complex arguments adduced in this
in span tasks, and memory span is more closely related to order paper, I would like to suggest a somewhat different way of think-
memory than item memory (e.g., Martin 1978), it is likely that the ing about the problem, based on an emerging understanding of
demands of the contextual component, rather than the number of the brain basis of consciousness (e.g., Baars 1993; 1998; Crick
activated items, set a limitation on capacity. 1984; Damasio 1989; Destexhe et al. 1999; Engel et al. 1999; Hob-
Against magical numbers. It has long been known that in dot son 1997; Newman et al. 1997; Steriade 1993; Tononi & Edelman
enumeration studies, RT increases linearly as the display size 1998). The magic number 7 plus or minus 2, which Cowan per-
increases above 4 items, suggesting that some serial counting suasively maintains is really about half that size, has become so fa-
process is taking place. One account of the processing disconti- miliar to scientists that we rarely pause to think how extraordinar-
nuity is that the error in number estimation increases systemati- ily small it is, given a brain with hundreds of billions of neurons,
cally with array size (Vos 1982). Hence, there is following Webers each firing at 10 Hz or faster, and each densely connected to all
Law (Van Oeffelen & Vos 1982). Hence, there is no discontinuity the others in only a few steps.
in number estimation between small values within the subitizing Limited capacity, which is always associated with consciousness,
range, and larger values. Rather, when estimating small numbers presents a great paradox: How could it make sense for it to be so
of discrete items, the estimation error is sufficiently small to allow small? Humans and animals must often run into danger because
a response of the nearest corresponding integer. With larger num- of their tiny capacity for immediately retrievable information. It is
bers, the error range may encompass several integers, and thus to simply implausible to think that a larger memory capacity could
maintain accurate performance at larger array sizes requires re- not evolve over hundreds of millions of years of brain evolution.
cruitment of a new algorithm, counting. To treat error-prone per- When a giraffe bends down to drink at a water hole, it cannot at
formance as an indication that certain items have not been the same time keep track of its young, pay attention to possible
processed results in the imposition of an artefactual capacity limit. predators, check for competing giraffes in the herds dominance
This concern haunts all models which propose a fixed integer hierarchy, and see if some sexual competitor is making eyes at its
number of slots, pointers or entities which govern STM capacity mate. If we add the need to make choices among action alterna-
(e.g., Conrad 1965; Trick & Pylyshyn 1994). Capacity is clearly tives, which also loads limited capacity, real-life situations like this
limited, because increasing the information load degrades perfor- must be informationally overwhelming at times. Being so easily
mance. But this does not imply dedicated mechanisms in 1:1 cor- overwhelmed with information must reduce our ability to respond
respondence with small arrays. rapidly and effectively to predators and other dangers. Like sleep,
Continuity of STM and LTM. This argument can be extended, which also exposes animals to danger, limited capacity is a biolog-
given the case made above, that the limitation of span-type stud- ical puzzle, because it appears to increase the risk to survival and
ies arises from temporal and spatial ordering of items. The preci- reproductive fitness.
sion with which discrete values can be classified on any dimension Indeed, the capacity for unrelated items in focal consciousness
is limited, as was demonstrated by many of the early psychologi- is even smaller than the size of immediate memory. I have sug-
cal applications of information theory. gested that conscious capacity is limited to only one single inter-
With small numbers of categories, classification is precise, but nally consistent event at any given instant one coherent chunk
as the categories increase in number, errors will occur. There is no at a time (Baars 1988; 1998). The evidence comes from numerous
reason here to consider that there is an abrupt discontinuity, and studies of ambiguous figures and words, in which only one in-
that beyond some limit, events are discarded from STM. An al- terpretation can be conscious at any instant. This is bolstered by
ternative conception proposes that over small ranges, ordering is evidence from dual-input tasks such as dichotic listening and
precise and retrieval rapid, whereas, with longer lists temporal binocular rivalry, and more recent experimental paradigms like
precision breaks down, and alternative encoding and retrieval pro- inattentional blindness (see Baars & Newman, in press). Other
cesses are employed. If the requirement for temporal precision is factors like visual scene complexity, which involves scenes that are
relaxed, for example, as in the studies of Jahnke et al. (1989), then hard to organize into coherent conscious units, also degrade our
performance improves. This could not occur if items were dis- ability to detect and respond to events.
carded from lists exceeding some arbitrary limit. If this position is It should be noted that very few items in short term memory
accepted, then the need to define a rigid partition between STM are conscious at any moment. Immediate memory has both con-

scious and unconscious components, as becomes obvious simply an answer. First, the primary function of the limited capacity com-
by considering how many items are reportable at any given instant ponent associated with consciousness is to allow coordination, in-
as conscious. In a rehearsed set of unrelated items, the Sternberg tegration, and global access between elements of the massively
task suggests that only one can be retrieved and reported as con- parallel-distributed society of neural nets. Second, I would sug-
scious at any given moment (Sternberg 1966). Thus, the momen- gest that a crucial aspect of the problem is the way in which con-
tary capacity of consciousness may be even smaller than that of scious cues may be used to index and retrieve unconscious chunks
immediate memory. It can be argued that immediate memory of information in immediate memory. Only within such a system
rides on the function of consciousness in the brain. Both have architecture can we begin to make sense of the radically small size
radical limits, in sharp contrast to such things as longterm episodic of immediate memory in a brain of extraordinary size, complexity,
memory or linguistic knowledge. In humans, immediate memory and moment-to-moment adaptive flexibility.
is enhanced by internal rehearsal and the visual sketchpad (Bad-
deley 1993). But the apparatus of consciousness is not limited to
humans; indeed, the most basic substrate of conscious states may
exist in all vertebrates, and certainly in mammals. The magical number 4 5 7: Span theory
Such a built-in capacity limit suggests a biological tradeoff. In
exchange for this radical limit, vertebrate species must gain some on capacity limitations
compensating advantage. Such trade-offs are routine in living or-
Bruce L. Bachelder
ganisms and engineered machines alike. But what could be the ad-
Psychological and Educational Service, Morganton, NC 28655-3729.
vantage? The only plausible answer, it seems to me, is to consider
brucebachelder@hci.net
the brain as a system architecture, a society of vast numbers of
complex, specialized neurons, which cluster in anatomical
Abstract: According to span theory, a behavioral theory of the magical
columns, arrays, and pathways with both fixed and variable con- numbers, Cowans 4 and Millers 7 are simply two different points on the
nectivities controlled by neurochemical modulation (Hobson same ogive describing the relation between performance and span load, a
1997). The most obvious circadian modulation of systemwide con- fundamental task characteristic. Span theory explains the magical num-
nectivity has to do with the changes due to conscious waking, bers in terms of a unitary limited span ability, a mathematical abstraction
dreaming, and unconscious slow-wave sleep. During conscious from that ogive.
states, massive reentrant adaptive resonance takes place in the
thalamocortical core, yielding a distinctive electrical field signa- With this paper, cognitivist Cowan becomes something of a
ture (Destexhe et al. 1999; Steriade et al. 1993). strange bedfellow in efforts to assert span theory (Bachelder &
Wolf Singer et al. have made a persuasive case in recent years Denny 1977a; 1977b). Cowan is right, there is a magical number,
that conscious perceptual states are associated with synchronized it is unitary, and developmental and individual differences are im-
oscillatory activity in the 40 Hz range (e.g., Engle et al. 1999), while portant (sect. 4.1.4); but he is wrong (sect. 4.3.1) that Millers idea
others suggest that thalamocortical firing is shaped by an was mere rhetoric. Miller (1956) evaluated a unitary channel ca-
information-theoretic parameter called complexity (Tononi & pacity hypothesis but his analyses failed (Bachelder 2000). The no-
Edelman 1998). These integrative features of the massive dyna- tions of span ability and a capacity-limited focus of attention
mic core presumably underlie conscious limited capacity; but promise to fare better.
they also make possible global access between conscious contents Figure 1 plots performance in the magical number tasks as a
and unconscious processes involved in learning, memory, motor function of numbers of items in a stimulus set. The curve, an in-
control, executive processes, and the like. Further, the thalamo- verse ogive, suggests there is no fundamental difference between
cortical core shows remarkable moment-to-moment adaptive Cowans 4 and Millers 7; both are simply different points on the
changes as a function of attentionally selected incoming stimula- same curve.
tion. Thus consciousness combines limited focal contents with The similar performance limitations in these tasks are usually
massive brainwide adaptation to selected input. These general fea- presumed to be coincidental. Cowans notion of a capacity-limited
tures command wide agreement (see Baars & Newman, in press). focus of attention undermines that presumption (sect. 4.3.3). Sec-
What could be the functional basis for a narrow coordinative cation 3.3.2 extends the theory to span of apprehension (known at
pacity in a massively parallel society of specialized neuronal ar-
rays? A long theoretical tradition in psychology suggests an an-
swer. Cognitive architectures were first developed in the 1950s by
scientists including Alan Newell and Herbert A. Simon, and have
since been developed by many others, notably John R. Anderson
(see Baars 1988). All cognitive architectures combine a parallel-
interactive set of memory elements with a narrow-capacity chan-
nel, which allows the memory elements to interact and be coordi-
nated for the purpose of problem solving. Some artificial
intelligence researchers have suggested that such global work-
space systems provide a general purpose problem-solving archi-
tecture. When predictable problems are encountered, such an ar-
chitecture can rapidly retrieve a canned response; and when an
unpredictable situation must be faced the architecture allows
coalitions of specialized knowledge sources to generate potential
solutions. It is striking that all our integrative models of human
cognition have these basic features; A massive parallel-interactive
memory with a narrow-capacity channel used for interaction, co-
ordination and control. I have suggested that in fact the brain has
found a similar architectural solution to its distinctive biological Figure 1 (Bachelder). The probability of a correct response as a
challenges (Baars 1988; 1993; 1998). The dynamic activity of the function of size of stimulus set in the magical number tasks.
thalamocortical core described above may be the brain version of Millers 7 corresponds to a 50% criterion, Cowans 4 to a 100% cri-
a cognitive architecture. terion. Pollacks data have been corrected. Mean performances on
While these points do not directly yield a numerical estimate for single stimuli were raised to the power of span load to estimate
the size of immediate memory, they suggest a framework for such performance on a full stimulus set considered collectively.

one time as span of attention). Section 4.4 sketches an extension closely onto Cowans capacity-limited attentional system. In the
to absolute judgment which would complete a unitary account. original model, as in Cowans, the system was assumed to have
Span theory accounts for all three limits via a mathematical both attentional and storage capacity. This issue will be discussed
proposition, an empirical generalization from data such as shown below. An important difference however, is that Cowans model
in Figure 1: In any task the probability of a correct response is an concerns short-term memory (STM), and as such is concerned
inverse ogival function of span load. Span load is the number of with explaining how people remember small amounts of material
stimuli jointly relevant for the target response. Span ability, oper- over brief time intervals, while the working memory model tries
ationalized as the 500 threshold of the Performance X Span Load to emphasise the functional role of a complex memory system
ogive, is the ability to cope with span load. This definition closely across a wide range of cognitive tasks.
parallels Millers definition of channel capacity as the upper limit It is probably this narrower focus that has caused Cowan to con-
on the extent to which the observer can match his responses to the centrate in his contrasts between the two systems on one compo-
stimuli we give him (p. 82). nent of working memory, the phonological loop. The loop does in-
Span theory analyses may help Cowan avoid a cul-de-sac in an deed, as Cowan emphasizes, have a time-limited component, the
extension of his theory to uni- and multidimensional absolute phonological store. Cowan does, however, appear to ignore the
judgment. First, Figure 1 shows that at Cowans 100% criterion overall attentional control system, the CE, the component that re-
the magical number in absolute judgment is 4, not 7. Second, the sembles his limited capacity model most closely.
principle that increasing the number of dimensions overcomes the Cowan himself assumes a range of other systems, which need
magical number may be an artifact of the information metric. to incorporate trace decay or interference, but dismisses those of
Miller relied largely on a study by Pollack and Ficks (1954). Bach- our own model as dubious on the grounds that one can always
elder (1978, Part 2) translated their data from information terms find examples that dont fit in giving as an instance spatial infor-
back to probability of a correct response, then modeled their task mation conveyed through acoustic stimulation. Given that the
as a combination of serial recall and unidimensional absolute judg- core studies on which the concept of a visuo-spatial sketchpad
ment. Subjects judged the values in each of multiple dimensions were based involved the auditory presentation of visuo-spatially
then retained and reported the values as in serial recall. Mathe- codeable material (Baddeley et al. 1975), this is somewhat sur-
matical models, presuming unitary magical number limitations prising. It presumably stems from his confounding the sketchpad,
typical for college students, generate the published data accu- an integrative system which explicitly combines visual and spatial
rately (mean error 5 2.7%). information from a range of modalities, with modality-specific
Span ability is a g-like trait construct. Figure 1 presents just one sensory memory. Cowans claim that the components of our model
of a family of parallel curves starting at the left for the develop- are situation-specific and therefore arbitrary is also puzzling,
mentally young (retarded and nonretarded). Memory span and given the extensive use of converging operations to define the
span of absolute judgment have been found to correlate .78 (60 working memory model, and its very wide application outside the
adults: college students, average, and retarded [IQ 5 47.8]; sum- memory laboratory (Baddeley & Logie 1999; Gathercole & Bad-
marized in Bachelder & Denny 1977a, pp. 13942). All three deley 1993).
magical numbers of mildly retarded adults are smaller (5 6 2) than Cowans claim that apart from his attentional control system no
those of college students (Spitz 1973). other mental faculties are capacity limited is also puzzling. Can
Cowans concepts map to some extent onto span theory. A ca- he really mean that every other cognitive system has infinite ca-
pacity-limited focus of attention is not that different from the no- pacity? Surely not. Perhaps he means that they are limited by fac-
tion of span ability, which can be characterized as the ability to at- tors other than attentional capacity. However, if one takes this
tend to several stimuli simultaneously. Chunks usually correspond claim with a pinch of salt, and identifies Cowans limited capacity
to responses. Recoding corresponds to changes in response reper- system with the central executive, then his model resembles the
toire. 1974 working memory model, with under-specified slave systems.
Span theory tightly integrates psychometric concepts and meth- As Miyake and Shah (1999) suggest, most current models of work-
ods into S-R style analyses of cognitive tasks. A true competition ing memory, whether explicit, or like Cowans, implicit, have a
of cognitive and behavioral approaches in the analysis of the same great deal in common. There are however, differences between
tasks promises to enrich both traditions. my own current model, and that implied by Cowan, including:
1. Cowan, in common with a number of other theorists, regards
short-term memory as simply the currently activated areas of long-
term memory. While this proposal has the advantage of simplicity,
it is inconsistent with neuropsychological data. Deficits in long-
The magic number and the episodic buffer term memory are found which have no apparent impact on im-
mediate memory, or on working memory more generally (Wilson
Alan Baddeley & Baddeley 1988), while both patient data and neuroradiological
Department of Experimental Psychology, University of Bristol, Bristol evidence indicator-specific short-term storage capacities (Basso et
BS8 1TN, United Kingdom. alan.baddeley@bristol.ac.uk al. 1982; Smith & Jonides 1995).
www.psychology.psy.bris.ac.uk/psybris 2. Conceptualization of Cowans limited capacity system differs
somewhat from our view of the CE, although both are relatively
Abstract: Cowans revisiting of the magic number is very timely and the loosely specified. Cowan appears to regard his attentional system
case he makes for a more moderate number than seven is persuasive. It is as unitary, involved in most aspects of memory, and anatomically
also appropriate to frame his case within a theoretical context, since this
will influence what evidence to include and how to interpret it. He pre-
as depending principally on the parietal region. I regard the CE
sents his model however, as a contrast to the working memory model of as a fractionable system that is less involved in retrieval than in en-
Baddeley (1986). I suggest that this reflects a misinterpretation of our coding (Baddeley 1996; 1999; Baddeley et al. 1985; Craik et al.
model resulting in a danger of focusing attention on pseudo-problems 1996). I also assume that it depends principally on the frontal
rather than genuine disparities between his approach and my own. lobes (Baddeley & Logie 1999).
3. Cowan seems to assume that his attentional system has stor-
In proposing their model of working memory, Baddeley and Hitch age capacity, otherwise, how are the sources of information inte-
split the earlier concept of the unitary short-term memory into a grated and maintained? While this assumption was made by our
tripartite system: a sound or language-based phonological loop, a 1974 model, it was later abandoned, with the CE regarded as a
visuo-spatial sketchpad, and an attentionally limited controlling control system using storage from elsewhere in WM or LTM (Bad-
system, the central executive (Baddeley 1986; Baddeley & Hitch deley & Logie 1999). Very recently, I have become convinced of
1974). Within this model, the central executive (CE) maps most the value of postulating a fourth component of working memory,

the episodic buffer which serves specifically as a limited capacity of items into larger chunks? The emergence of the chunk size will
store in which information can be integrated from the slave sys- be identifiable by performance discontinuities and various indi-
tems and long-term memory and represented in a multi-feature rect effects of the capacity limits (Cowans final two basic condi-
code. As in Cowans model, this is assumed to have a limited cations). Performance discontinuities and other effects of a limited
pacity determined by number of chunks, and to involve a retrieval capacity should then be predictable on the basis of the ways in
component that is associated with conscious awareness. In addi- which recoding has been impeded.
tion to serving as a store, it is assumed to provide the modelling What is clear from this analysis is that Cowans argument would
space where information can be combined and manipulated in or- benefit from a stricter definition of how to prevent higher level re-
der to plan future actions or interpret recollected experience coding than he currently provides. This may also require further
(Baddeley 2000). It is thus assumed to be a temporary memory work in investigating the nature and limitations of the processes
storage system. As such, it differs from the CE, an attentional con- contributing to compound STM limits in order to definitively rule
trol system that operates across many tasks other than memory. out their involvement.
The episodic buffer is assumed to combine and store information The second way of testing Cowans argument is, at first blush,
from different sources, allowing active manipulation so as to cre- the simplest. In addition to studies cited by Cowan there are nu-
ate new representations that serve to solve novel problems. As merous others in the literature that address the question of pro-
such it provides the temporary storage needed for mental model- cessing limitations. Any that conform to Cowans basic conditions
ling. but do not demonstrate signs of a 4-chunk capacity limitation will,
I therefore welcome Cowans review of evidence relating to the in effect, falsify Cowans thesis. Once again, there is the problem
limited capacity of immediate memory. I prefer to attribute it of ensuring that high-level recoding has been adequately impeded
however to the limited capacity of the central executive and its re- but if a procedure is available that impeded recoding in an earlier
lated storage system, the episodic buffer. experiment it is reasonable to assume that recoding was equally
well impeded even if measured capacity exceeds 4 chunks.
Finally, consider the following intriguing scenario suggested by
Cowan in relation to scene coherence. Given a 4-chunk capacity-
limited STM we must assume that scenes may have up to 4 sep-
The size and nature of a chunk arate parts, or events, in awareness at any given moment that,
although associated with a common higher-level node, do not nec-
C. Philip Beaman
essarily have any common level of association beyond this. To con-
Department of Psychology, University of Reading, Whiteknights, Reading
nect them will require shifting attention from one event to another
RG6 6AL,United Kingdom. c.p.beaman@reading.ac.uk
www.rdg.ac.uk/acadepts/sx/Psych/people/beaman.html
in a manner sub-optimal for later serial recall. Within the auditory
domain experiments of this nature have in fact been conducted.
Abstract: The data presented in the target article make a persuasive case
Performance is worse when untrained listeners are required to re-
for the notion that there is a fundamental limit on short term memory port the order of a recycled sequence of four unrelated sounds
(STM) of about four items. Two possible means of further testing this than when the sequence contains sounds of a similar type (War-
claim are suggested and data regarding scene coherence and memory ca- ren & Obusek 1972; Warren et al. 1969). These studies suggest
pacity for ordered information are reviewed. that order information is indeed impaired when items are not as-
sociated.
Given the current state of knowledge, it is difficult (but not im- A further study from a different domain, that of thought sup-
possible) to argue against the idea that there is some fundamen- pression (Wegner et al. 1996) also concludes that memory for or-
tal, architectural constraint on processing capacity (Just et al. der requires some level of association. Wegner et al. found that if
1996; Young 1996; Young & Lewis 1999). It may be even harder, participants were required to suppress memory for the sequence
however, to point to exactly what that constraint is. The com- of events in a film sequence, information was lost but not item in-
pound STM limits Cowan refers to in the target article reflect formation (as measured by free and cued recall and by recogni-
functional rather than architectural constraints on what consti- tion). These studies, although they do not speak directly to the is-
tutes processing capacity, leaving Cowan with the unenviable job sue of capacity, confirm Cowans intuition that shifts of attention
of defining the fundamental pure STM capacity limit in terms between different events are necessary to set up sequence infor-
of chunks of information. As with Millers classic article on the mation for events which otherwise are not associated. Such se-
same topic, however (Miller 1956), there is no satisfactory, inde- quence information can be destroyed by active attempts to sup-
pendent definition of a chunk. press the memory of the events. The studies, however, also raise
The circularity of the logic becomes apparent thus: chunks can questions about the nature of an association. If the processing
be identified with individual stimulus items when information limit on STM is 4 chunks, what makes information about the or-
overload or other experimentally induced conditions prevent cod- der of those chunks easier or harder to recall? The question is one
ing of the stimulus items into larger or, arguably, higher-level that is of more importance than might immediately appear. A
chunks. How can one identify a chunk in any particular situation? number of studies which Cowan takes as evidence for his 4-chunk
When steps have been taken to prevent coding of an item to a capacity limit were serial recall studies in which items were only
higher level, the item will be the chunk. How do we know whether marked correct if they were recalled in the correct serial order
the experimental conditions have been successful at preventing (e.g., Table 2 in target article). It seems therefore that the capac-
recoding? Because the item will act like a chunk. In other words ity limit reflects not a limit on capacity to recall individual items
an item will act as a chunk when it has not been recoded. We know but a combined limit on recall of the items and the (order) rela-
it has not been recoded because it is acting like a chunk. tions between them. In serial recall, correct recall of items is in-
In fact, Cowan provides a way round this problem by his analy- terdependent. If an item is misrecalled early in the list it is unlikely
sis of the conditions under which a chunk size of 4 emerges, his to be correctly recalled later in the list as well as blocking recall of
four basic conditions. His argument therefore stands. As an argu- the correct item at the earlier position. If capacity as measured by
ment. It is, and should be, open to empirical test. There appear to serial recall studies is indeed 4 chunks, then capacity is for 4
be two obvious ways of going about this. The first way is to imple- chunks plus some extra information connecting those chunks. It is
ment the constraint in some formal model that also includes the then necessary to determine the nature of this extra information
four basic conditions for identifying a chunk. Are chunk sizes of 4 and its relation to chunk capacity.
really observed when information overload in some (as yet un-
specified) way prevents higher level recoding in simulation stud-
ies? Or when other steps are taken specifically to block the coding

There is no four-object limit on attention many previous studies of visual attention. In each trial, the six or
three objects were initially presented for 2.4 seconds to give sub-
Greg Davis jects ample time to focus their attention on the objects. Next, two
Department of Psychology, Birkbeck College, London WCIE 7HX, United notches were removed from the objects in the display. The
Kingdom. g.davis@psyc.bbk.ac.uk www.psyc.bbk.ac.uk/staff/gid.html notches were either square or jagged and the position of one notch
in no way predicted the position of the second notch, with every
Abstract: The complex relationship between attention and STM forms a possible combination of notches and notch-positions being
core issue in the study of human cognition, and Cowans target article at- equiprobable. The task was simply to determine whether the two
tempts, quite successfully, to elucidate an important part of this relation- notches were of the same type (i.e., both jagged or both square)
ship. However, while I agree that aspects of STM performance may reflect or of different types (i.e., one jagged, one square), and to press the
the action mechanisms that we normally consider to subserve attention
I shall argue here that attention is not subject to a fixed four-object capacity
appropriate key on a computer keyboard as quickly as possible.
limit as Cowan suggests. Rather, performance in attention tasks as well as Performance was indistinguishable for the six- versus three-
STM may be best accounted for in terms of decay and interference. object displays, suggesting that no four-object limit had operated,
and that six objects can be attended as efficiently as three objects,
Many previous studies have concluded that observers can track or once other display variables are equated in the two cases. How-
enumerate up to four small items efficiently, but cannot do so for ever, one valid objection to this new study might be that observers
more than five objects, this restriction reflecting a four-object had not perceived the six-object displays to contain six separate
limit on our visual attention (e.g., Pylyshyn 1989; Trick & Pylyshyn objects. Rather they may have perceived each pair of objects in
1993). However, these studies have largely employed displays those displays to comprise a single object or Gestalt, and thus per-
such as those in Figures 1A and 1B, where several other display ceived only three objects in the six-objects displays. To preclude
variables co-vary with the number of relevant objects. For exam- this possibility, Davis et al. examined performance on some spe-
ple, the five objects in Figure 1B constitute a greater overall sur- cific trials within the three- and six-object displays. For six-object
face area than do three objects in Figure 1A, and presumably com- displays they compared RTs when the two features by chance ap-
prise a greater number of visual features overall. Accordingly, peared on a single small object (i.e., were horizontally-displaced
any costs for attending more versus fewer objects in such displays from each other) versus appearing on two separate neighboring
may not reflect the number of relevant objects per se, but might objects (vertically-separated within a single pair of objects as in
alternatively result from the greater amount of relevant percep- Fig. 1D). For three-object trials they conducted an identical com-
tual information comprised by larger numbers of objects. parison of horizontally- versus vertically-displaced feature-pairs,
In order to remove this ambiguity in their own study, Davis et except that the features now always appeared upon a single large
al. (submitted) have compared attention to displays of three large object.
objects (Fig. 1C) versus six small objects (Fig. 1D), in which only Many previous studies have demonstrated that comparison of
the number of objects varied. Note that although the two display features within pairs of objects is more efficient when they appear
types comprise different numbers of objects, they are otherwise on the same object than on two separate objects (e.g., Davis 2000;
very much alike, so that the six objects comprise approximately the Duncan 1984; Lavie & Driver 1996; Watson & Kramer 1999).
same surface area and number of features as do the three ob- Thus if the six-object displays had indeed been perceived to com-
jects. If there is indeed a fixed four-object limit on attention, then prise six objects, and the three-object displays, three objects, the
attending the six-object displays should entail performance costs following patterns of performance should be expected. First, hor-
relative to attending the three-object displays. Conversely, if ap- izontally separated features (on the same small object) in the six-
parent four-object limits in previous studies reflected the overall object displays should be responded to faster than vertically-
surface area and number of features that co-varied with the num- separated features in those displays (lying on different, neighbor-
ber of objects there, the two displays should now yield equivalent ing objects). Second, if these differences reflected whether the
performance. two features were on same versus different objects, rather than
To measure how efficiently the two displays could be attended, simply resulting from other differences between vertically and
Davis et al. employed a divided-attention task similar to those in horizontally-separated features, this pattern of results should not
hold in the single large objects of the three-object displays. Pre-
cisely this predicted pattern of results was found, validating Davis
and colleagues comparison of six versus three objects.
Accordingly, Davis et al. suggested that no four-object limit
holds for visual attention. Indeed, attention may not be limited to
any magica1 number of objects, but rather may reflect the same
properties of decay and interference that Cowan ascribes to STM.
Recent evidence points to the existence of two distinct binding
mechanisms in vision: within-object links that bind together fea-
tures from the same perceptual objects and between-object links
that bind features from separate objects. These two types of link
appear to be coded in anatomically separate visual streams, so
that there may be substantial mutual interference between links
of the same type, but very little interference between the two
types of link (see Humphreys 1998). Davis et al. suggested that
this possible mutual interference between links might explain
both the equivalent performance found for their six- versus three-
object displays, and a range of other findings previously ascribed
to a four-object limit. First, since their six- versus three-object dis-
plays comprised approximately equal numbers of features, pre-
sumably bound by the same number of links, interference should
have been approximately the same in the two cases, correctly pre-
dicting the equivalent performance found. However, in more con-
ventional displays, such as those in Figures 1A and 1B, where the
number of features increases with the number of objects, mutual
Figure 1 (Davis). interference between links should increase, decreasing the effi-

ciency with which the objects can be attended. In fact, when Davis tions formed between presented chunks. More recently, Ericsson
et al. calculated how within and between-object links might be ex- et al. (2000) showed how LTWM mechanisms developed to sup-
pected to increase for 2 4 objects versus for 5 7 objects, this port expert performance can transfer and lead to superior perfor-
yielded expected performance decrements of 37 and 235 msec/ mance in traditional memory tasks involving domain-related in-
item-1 respectively, closely approximating previously reported formation.
values of 40 and 240 msec/item-1. Clearly, this very simple ac- Cowan acknowledges that traditional STM tests can be influ-
count models only performance rather than underlying cognitive enced by LTM storage and techniques such as rehearsal, and sug-
processes. However, it does provide an existence proof that ap- gests that this type of performance yields compound, not pure
parent four-object limits in previous studies can in principle be ac- estimates of STM capacity. He claims that it is possible to experi-
counted for in terms of interference alone. mentally control for these contaminating influences, using exper-
I conclude that there is no four-object limit on attention, and imental procedures that have been designed to prevent partici-
that the Davis et al. view of attention limitations may generate far pants from rehearsing and encoding presented items and their
more biologically plausible models than accounts that assume a associations in LTM. Cowan argues that memory testing with
magic number of objects. However, the absence of a four-object these experimental procedures will uncover pure STM with a uni-
limit on attention does not invalidate attempts to ascribe aspects versal capacity of four independent chunks.
of STM performance to attention processes. Indeed, a common We question the effectiveness of procedures advocated by
interference-based framework for understanding STM and atten- Cowan to eliminate storage in LTM and restrict recall to inde-
tion may assist such efforts. pendent chunks from pure STM. Many of the prototypical pure
STM studies Cowan cites in Table 2 involve recall of meaningful
stimuli, such as unrelated words, presented while subjects are en-
gaged in rehearsal suppression. To test how effectively these pro-
The search for fixed generalizable limits cedures control encoding, we examined how skilled and expert
of pure STM capacity: Problems participants performed under such conditions. In contrast to the
studies cited by Cowan, rehearsal suppression during encoding
with theoretical proposals based had little or no effect on skilled and expert performers. They were
on independent chunks able to encode and to form associations between different pieces
of presented information in LTM (see Ericsson & Kintsch 1995
K. Anders Ericsson and Elizabeth P. Kirk for a review), even when the information was auditorily presented
Department of Psychology, Florida State University, Tallahassee, FL (Chase & Ericsson 1981; Ericsson & Folson 1988).
32306-1270. {ericsson; kirk}@psy.fsu.edu Another approach proposed by Cowan for minimizing storage
www.psy.fsu.edu/~ericsson
in LTM and assessment of pure STM involves examining memory
for presented material that cannot be easily be rehearsed. How-
Abstract: Cowans experimental techniques cannot constrain subjects re- ever, studies of expert performance are replete with examples of
call of presented information to distinct independent chunks in short-term
memory (STM). The encoding of associations in long-term memory con-
experts who exhibit superior recall for information that cannot be
taminates recall of pure STM capacity. Even in task environments where easily rehearsed, such as chess positions.
the functional independence of chunks is convincingly demonstrated, in- The ability to rapidly encode and integrate information in LTM
dividuals can increase the storage of independent chunks with deliberate is not restricted to skilled performers and experts in their domains
practice well above the magical number four. of expertise. During the initial trials of memory testing (prior to
the build-up of practice interference), storage in LTM is consid-
Cowan is continuing the search for a magical number that mea- erable even with typical stimuli and college students (Ericsson &
sures the universal limit for the pure capacity-based STM (pure Kintsch 1995). With more sensitive recognition based techniques
STM). Miller (1956) first proposed the magical number seven to relational storage in LTM is observed even during repeated test-
describe the remarkable invariance in performance on laboratory ing ( Jiang et al. 2000). It is doubtful whether Cowans recom-
tests of short-term memory (STM). He discovered that individu- mended procedure can reliably prevent relational encoding of
als encoded stimuli as familiar distinct patterns (chunks) already presented information with storage in LTM during a memory test
stored in long-term memory (LTM) and could recall roughly the and thus indirectly guarantee independence of recalled chunks.
same number of chunks regardless of type of presented items, To rigorously test the validity of Cowans fixed chunk limit of
such as letters or words. When no associations are formed be- pure STM, we must identify task environments in which suc-
tween chunks each chunk must reside independently in transient cessful performance requires independence of chunks. When
STM to be accurately recalled. Consequently, Miller proposed tasks prevent individuals, including highly skilled performers,
that the storage capacity of STM had an invariant general limit that from relying on associations between chunks, the evidence is in-
could be expressed as approximately seven independent chunks. consistent with Cowans fixed limit of four independent chunks.
Subsequent research has shown that the fixed capacity limit of For example, during mental addition of long sequences of num-
independent chunks does not constrain STM after extended prac- bers individuals must maintain the running sum as individual dig-
tice (Chase & Ericsson 1982; Richman 1995), and for skilled and its that can be updated and changed independently of each other.
expert performance (Ericsson & Kintsch 1995; Gobet & Simon Skilled abacus operators hold running sums of three or four digits
1996). In domain-specific tasks, experts are able to encode and while performing mental computation. This number of indepen-
store presented information in LTM and expand the capacity of dent chunks is consistent with Cowans estimate of pure STM
working memory (WM). yet when operators engage in deliberate practice they can increase
Ericsson and Kintsch (1995) showed that skilled individuals can the number of digits of the running sum by one digit per year
store information in long-term working memory (LTWM) by in- (Stigler 1984). Skilled abacus operators can expand their memory
tegrating two encoding mechanisms. First, they are able to gen- performance to 14 16 digits with minimal disruption by rehearsal
erate associations between different presented chunks of infor- suppression or preloading STM with unrelated information (Ha-
mation and build new integrated structures in LTM. If experts can tano & Dsawa 1983). They maintain efficient access to indepen-
encode associative relations between virtually all chunks within dent digits and can recall up to 10 digits equally fast in forward or
their domain of expertise, then the concept of chunk indepen- backward order. This large expansion of the functional capacity is
dence would not apply. Second, and more important, skilled inconsistent with the second type of acquired mechanism of LTWM
dividuals acquire skills to associate chunks of presented infor- that involves recency-based encodings (Ericsson & Kintsch 1995).
mation with internal cues organized in retrieval structures. LTM Expert performers ability to modify their functional pure
storage mediated by this mechanism would not require associa- STM capacity in a specific domain raises issues about the gener-

alizability of fixed limits of capacity measured by independent limitation may result from a general property of information
chunks. And our point is not restricted to experts. Expert perfor- processing systems related to the computational requirements
mance results from years of gradual skill acquisition (Ericsson needed to achieve distinguishable mental representations of ob-
1996). Similar types of LTWM mechanisms, albeit in less refined jects (computational view). According to this view, whereas each
form, mediate the extended development of many types of skills. hemisphere may independently have a memory capacity of four,
Ericsson and Kintsch (1995) showed that the same type of LTWM the combined memory capacity across the two hemispheres would
mechanisms mediate reading and other everyday activities medi- still be four (at least in normal, neurologically intact young adults).
ated by comprehension. When individuals perform tasks involv- These three views of working memory (which are neither mu-
ing comprehension, they encode relevant, associatively related in- tually exclusive nor exhaustive) lead to different predictions about
formation to guide their performance. Associative encoding and what would happen if a working memory task involved presenting
integration of encountered information is an essential part of this the information exclusively to one or the other hemifield (and
process. If most cognitive activities in ecologically valid situations therefore to one hemisphere) or distributing it across both hemi-
do not involve storage of independent chunks, it is unlikely that fields (thus dividing the information between the hemispheres).
pure STM capacity limits based on independent chunks will be Specifically, the distributed view would predict an increase in ca-
relevant predictors of performance. pacity when information is distributed across the two hemifields,
whereas the unitary and computational views would not. In addi-
tion, large differences in capacity between left and right visual
field presentations would be more consistent with the unitary than
with the computational view.
Working memory capacity and the For this logic to be applicable it is important to be able to ma-
hemispheric organization of the brain nipulate the hemisphere in which the information is activated
and/or maintained. Various approaches to this problem can be
Gabriele Gratton,a Monica Fabiani,a considered, including the type of material to be maintained (e.g.,
and Paul M. Corballisb verbal vs. spatial) and the location in space where the stimulus is
a
Department of Psychology, University of Missouri-Columbia, Columbia, MO presented (e.g., left vs. right visual fielddivided field paradigm).
65203; bCenter for Cognitive Neurosciences, Dartmouth College, Hanover, In a series of recent studies (Fabiani, in press; Gratton et al. 1998)
NH 03755. {grattong; fabianim}@missouri.edu we have obtained evidence for the hemispheric organization of vi-
paul.m.corballis@dartmouth.edu sual memory. The basic paradigm used in this work involved: (1)
www.missouri.edu/{~psygg; ~psymf}
the presentation of information to one visual hemifield at study
(and therefore, at least initially, to one hemisphere of the brain);
Abstract: Different hypotheses about the mechanisms underlying work- and (2) the subsequent testing of memory for the item when it is
ing memory lead to different predictions about working memory capacity
when information is distributed across the two hemispheres. We present
presented at different locations within the same or the opposite
preliminary data suggesting that memory scanning time (a parameter of- hemifield. Typically, in these studies we observed a reduction in
ten associated with working memory capacity) varies depending on how performance when the stimulus was presented in a different
information is subdivided across hemispheres. The data are consistent hemifield during study and test, compared to conditions in which
with a distributed model of working memory. the stimulus was presented within the same hemifield during both
study and test. Further, if the stimulus was presented centrally
The Cowan target article emphasizes the limited capacity of work- (and therefore to both hemispheres) at test, the brain activity
ing memory, and presents a varied and significant body of evi- elicited during the test phase was systematically lateralized (i.e.,
dence indicating that working memory capacity is approximately larger on the left or on the right) depending on the hemifield
four items. Over the last two decades, researchers have become where the stimulus had been studied.
increasingly interested in understanding the relationship between These data suggest that the divided field paradigm can be used
cognitive function and the brain. A way to begin addressing the is-
sue of what type of brain mechanisms underlie the fundamental
limitation of working memory capacity emphasized by Cowan and
previous investigators (e.g., Miller 1956) is to consider the inter-
action between working memory capacity and the hemispheric
organization of the brain.
This interaction may take different forms. According to one idea
working memory is a single, unified resource, or module. This
suggests that working memory may be localized in one hemi-
sphere only (for instance, in the left hemisphere) rather than be-
ing distributed across the two hemispheres (unitary view). If this
were the case, distributing information to be held in working mem-
ory across the two hemispheres would not be advantageous, and
may in fact be deleterious in some cases. A variant of this view is
that a number of working memory systems each specialized for
different types of materials coexist, and each of them is localized
(absolutely or relatively) in one hemisphere (e.g., left hemisphere
for verbal material and right hemisphere for spatial material).
A different idea is that working memory can be envisioned as a
distributed system (distributed view). In this case, the circuit
supporting working memory function could be partly imple-
mented in one hemisphere and partly in the other, even for the Figure 1 (Gratton et al.). Average reaction times for correct pos-
same type of stimulus material. Hence, distributing information itive trials in the divided visual field memory-search paradigm, as
across the two hemispheres may facilitate its activation and/or a function of the distribution of memory set items across the left
maintenance, leading to an increase in working memory capacity. and right visual field (23: two items in one hemifield and three
Finally, as indicated by Cowan, the limitation of working mem- in the other; 14: one item in one hemifield and four in the other;
ory capacity may not be due to a specific limitation in the mecha- 05: zero items in one hemifield and five in the other). Error bars
nisms used to activate and/or maintain information. Rather, this indicate the standard error of the mean for each condition.

to manipulate the hemisphere that holds information to be main- are looking for a magical number, an integer, perhaps because we
tained in working memory. In a pilot study, we applied this logic actually have a simple mind that imposes simplicity on the under-
to the question of whether distributing information across the two standing of phenomena, and also perhaps because it would be
hemifields may improve working memory performance. We used more useful to work on firm ground rather than rely on the capri-
a variant of the memory search paradigm (Stemberg 1966) in cious fluctuations of malleable chunk sizes or content.
which the memory set stimuli were presented one item at a time, We may not have a definite magical number yet, but we do
either to the left or to the right of a central fixation cross, in an un- have a magical principle to offer to answer the basic capacity
predictable fashion. The stimuli consisted of patterns of randomly question, which is: How can the connection between what is
oriented lines to minimize verbal rehearsal. Each item was pre- made available by the sensory systems and the organization of it
sented for 200 msec, with 800 msec ISI. by the brain (or higher processing structures) become as efficient
Following the presentation of the memory set, a test stimulus as possible in capturing the essence of available information?
was presented centrally. This could be one of the five memory-set Miller (1956) promoted this fundamental chunking/recoding
stimuli or a similar unstudied stimulus. The average reaction times principle: when . . . we want to remember, we usually try to
for the correct positive(yes) responses to test items are pre- rephrase [it] in our own words (p. 95). This is not without re-
sented in Figure 1. The data indicate a performance advantage calling notions such as the assimilation mechanism assumed by
(i.e., faster reaction time) when the memory set items were di- Piaget in the adaptation process; or the use of field theory in
vided across the two hemifields compared to those conditions in Gestalt, where the brains electrochemical forces, which underlie
which the memory set items were all presented to the same hemi- structured fields, transform sensory data and are transformed in
field. the simplest way possible.
In interpreting these data, it is important to note that the test Representations of mental capacity. Between what is already
stimuli were physically similar for the three conditions presented stored in the system and new inputs, it looks like something oc-
in Figure 1, and that the total memory set always consisted of five curs, and that an intermediate variable is needed to describe this
items. state. The name of this variable indicates the researchers stand.
Although preliminary, these data suggest that dividing informa- They call it memory short-term memory, in the target article
tion between the two hemifields may improve memory-scanning and the question that they most often address concerns its storage
time, which has been traditionally considered as related to work- capacity. In this context, a capacity hypothesis such as the dura-
ing memory capacity. These data are consistent with a distributed tion of an item in this memory without rehearsal may simply be ir-
view of working memory. relevant. The idea here is that terms like storage and capacity
often refer to the notion of volume, that is, to a spatial represen-
tation. Duration cannot express a notion having a spatial connota-
tion. If mental abilities or mental processes are the issue, then
A temporal account of the limited storage space is only one of a number of potential answers, and
a duration hypothesis which may or may not involve storage time
processing capacity can possibly be rehabilitated. It is indeed somewhat ironic to
discard a duration explanation from a short-term memory issue,
Simon Grondin
given that the word term itself expresses duration. This term
cole de psychologie, Universit Laval, Qubec GIK 7P4, Canada.
should, instead, be quantified, if not with a time unit, at least with
simon.grondin@psy.ulaval.ca www.ulaval.psy.ca
a descriptor of a fading rate.
Whether they are Working/Active/Immediate/Short-Term
Abstract: A temporal account of the mental capacities for processing in-
formation may not be relevant in a context where the goal is to search for
(WAIST?), all such varieties are in fact called memory. The word
storage capacity expressed in chunks. However, if mental capacity and in- memory makes sense if information is viewed as being retained
formation processing is the question, the time issue can be rehabilitated. or stored for some duration, rather than being present, available
A very different temporal viewpoint on capacity limit is proposed in this or circulating. However, although WAIST memory favours per-
commentary. manent storage, storing within WAIST memory is not the pur-
pose: processing is.
Using a reasonable set of boundary conditions for testing the In Cowans theoretical framework, the capacity issue enters a
chunk capacity, Cowan has helped distinguish between what be- new phase once the limit is said to depend on attention, more
longs to a pure storage estimate and what is a compound measure specifically, on the focus of attention. There are several ways of ex-
of storage. His impressive review of evidence defending the 4- pressing the involvement of attention in information processing.
chunk thesis comes from direct and indirect empirical facts in cog- Focus is one, but one that may also carry a spatial connotation,
nitive psychology, along with evidence from mathematical models that is, visuo-spatial attention, which, along with the storage ca-
and physiological data. Beyond the overall convincing support for pacity notion, serves to reinforce spatial representation.
the 4-chunk notion, Cowan returns the reader to the cause of this Attention limits are ultimately reported to be responsible for
capacity limit, which is reported to be attention. In the process producing chunk limits. If anything, chunking reveals more about
however, time-based interpretations were quickly discarded. It is attention than about memory. Chunking is the principle by which
argued below that the temporal option should and can be used for attention mechanisms are accommodated, and processing is made
describing the capacity limit. easier.
What are the magical chunk number and chunking for? We are Duration as information, rather than duration of information.
getting closer to a pure estimate of the capacity limit, as expressed Another vein in the literature of experimental psychology may en-
by the number of chunks. From 7 6 2 which was indeed 3 to 15, liven a debate on capacity limits, where considerations about du-
in Miller (1956), it passed to 4 6 1, which remains nevertheless 2 ration decay remain a persisting potential explanation. Because
to 6, in the target article. The numbers have changed and the mar- rehearsal provokes interference, either proactive or retroactive,
gin of error is narrower, but scientists eager to have a reliable law Cowan reports the decay issue to be unresolvable. Nevertheless,
or a constant for mental life to work with will probably remain un- this should not totally disregard attempts to provide a temporal ac-
satisfied. count of the state occurring at some point between sensory stim-
However, although the numbers have changed, the chunking no- ulation and permanent storage.
tion has not. Indeed, beyond any estimation of maximum chunks In the literature on time perception, there is a concept called
that can potentially be processed simultaneously (perhaps in very the psychological present for describing this state. This can be de-
rapid, but tractable, succession), we are dealing with a fundamen- fined as a highly flexible tuning process that is dynamically fitting
tal principle of mental processes: fragmenting information. We the temporal width of the field of attention and its phase relations

to the sequential structure of the pattern of events. (Michon Processing capacity limits are
1978, p. 89); or . . . the temporal extent of stimulations that can not explained by storage limits
be perceived at one given time, without the intervention of re-
hearsal during or after the stimulation (Fraisse 1978, p. 205). Graeme S. Halford,a Steven Phillips,b
These authors propose that the present has an average value of 2 and William H. Wilsonc
or 3 sec, with an upper limit of 5 (Fraisse 1978) or 7 or 8 sec (Mi- a
School of Psychology, University of Queensland, Brisbane 4072, Australia;
cho 1978). The wide variability of estimates is at least as disap- b
Information Science Division, Electrotechnical Laboratory, Tsukuba 305, Ja-
pointing as is that surrounding a maximum chunk number. Nev- pan; cDepartment of Computer Science and Engineering, University of New
ertheless, a 2-sec value is at least interesting in terms of recalling South Wales, Sydney, New South Wales, Australia. gsh@psy.uq.edu.au
the memory span estimate of Baddeley (1986). stevep@etl.go.jp billw@cse.unsw.edu.au
Just as for the variability in estimating maximum chunks, vari- www.psy.uq.edu.au/gshm www.cse.unsw.edu.au/~billw
ability in estimating the psychological present (in time units) de- www.etl.go.jp/etl/ninehi/stevep@etl.go.jp/welcome.html
pends on boundary conditions adopted. One way to look at the
problem is to refer to the discrimination law of Weber (Grondin, Abstract: Cowans review shows that a short-term memory limit of four
items is consistent with a wide range of phenomena in the field. However,
in press). This law states that, for one given sensory continuum,
he does not explain that limit, whereas an existing theory does offer an ex-
difference threshold (D) should increase as a function of the planation for capacity limitations. Furthermore, processing capacity limits
magnitude of the stimulus (), with D/ being a constant, k, the cannot be reduced to storage limits as Cowan claims.
Weber fraction. It is well-known that for very small sensory mag-
nitudes, this fraction is large and gradually becomes smaller, and
then stabilizes. In his excellent review, Cowan concludes that short-term memory
At this point, it is worth taking a look at Webers law for time storage is limited to four items, noting that this corresponds to the
perception, more specifically concerning duration discrimination limit in processing capacity defined by Halford et al. (1998). Fur-
under a condition where the only processing required is to judge thermore, his conclusion that the limit is in the number of inte-
the duration of a sensory signal or the empty duration between two grated objects, independent of the complexity of each, agrees well
brief sensory signals. For duration discrimination, the Weber frac- with the observation of Halford et al. (1998) that humans are lim-
tion presents a degree of constancy, but at some point, with longer ited to relating four entities, irrespective of their complexity. How-
intervals, some discontinuity also occurs (as in boundary condition ever, these correspondences do not imply that processing limits
3 in Cowans target article): the fraction becomes higher (see for can be subsumed under storage limits, as Cowan claims.
instance, Fig. 1 in Fraisse 1978). This discontinuity occurs at some The fact that the size of the limit is four in both cases is not a
point between 1 and 2 sec, and can be interpreted as a point mark- strong argument for identification because, given that the limit is
ing the upper limit of the psychological present span, that is, as small, the same number could occur in both contexts by coinci-
temporal factor accounting for capacity limit. dence. Alternatively, storage and processing systems could be dis-
Another operational definition can also be adopted for describ- tinct but with equal capacities to facilitate transfer from one to the
ing a critical limit in temporal processing. One strategy for reduc- other. There are a number of reasons why processing cannot be
ing difference threshold when the time intervals to be discrimi- subsumed under storage. To take a straightforward example, there
nated are long is to adopt a counting strategy, that is, using chunks. clearly is a difference between simply holding the numbers 7 and
One way of defining the upper limit of psychological present is to 4 in short term store, and adding them to yield the sum, 11. In gen-
look at the point from which chunking becomes a useful strategy. eral storage, in the sense of internal representation, is a prerequi-
This point can be estimated by examining two functions relating site for processing, but cognitive processing cannot be reduced to
difference threshold and time, one when counting (chunking) is storage. Furthermore higher cognitive processes require repre-
permitted, and one where subjects refrain from counting. Such sentations that have properties beyond those required for storage,
functions intersect at about 1.18 sec (Grondin et al. 1999). When including omni-directional access and analogical mapping (Hal-
intervals are longer than that value, using chunks is a more valu- ford et al. 1998).
able option; and this critical point can be viewed as an upper limit Cowans position is that a concurrent short term memory load
of the psychological present span. With optimal chunking, this can be held in the activated portion of long term memory while
value may be lower, but this remains to be demonstrated. A for- other information is being processed in the focus of attention.
mal analysis of the consequences of fragmenting intervals into Lack of interference between processing and short-term storage
subintervals, thus, of adopting chunking strategies in estimating is explained because the focus of attention can be devoted to ei-
time, is presented in Killeen and Weiss (1987). ther storage or processing, but need not be devoted to both at
Concluding remarks. Cowan has made a commendable contri- once. However this still implies that storage and processing are
bution to current investigations of processing capacity by con- distinct, and also implies there would be no tradeoff between the
necting multiple evidence of the 4-chunk storage. Nevertheless, two. It is not fundamentally different from the position of Halford
although the amount of information expressed in stored chunks is et al. (1998).
dimensionless, it may suggest a type of representation of mental Cowan offers no explanation for the limit he observes in stor-
capacity that is advanced so as to rule out viewing potential tem- age capacity, whereas Halford et al. (1998) offer a natural expla-
poral interpretations for processing capacity. Both views talk about nation for processing capacity limits. In this model, conceptual
a limited span of apprehension of reality or an active state of complexity is defined by the arity, or number of arguments that
consciousness. Confusion may arise from calling it a memory. can be bound into a single relation. Human adults are typically
Using terminology such as short-term memory, in opposition to limited to processing one quaternary relation in parallel. Each
long-term memory, is certainly an oversimplification of physical component of the relation is represented by a vector, and the bind-
time (term), which is a continuous variable. Information is in a ing is represented by the tensor product of the vectors. Thus, the
state varying from 0 to 100% of processing availability. Attention binary relational instance larger (elephant, mouse) is represented
mechanisms, which probably respond to sensory-mode require- by vlarger 3 velephant 3 vmouse. The rank of the tensor product is
ments and particularities, activate pieces of information for mak- one more than the arity of the relation. The more complex rela-
ing it available, present. Four such pieces of information can po- tions are represented by tensor products of higher rank, the
tentially become available without impairing performance. greater complexity of which explains why more complex relations
are associated with higher processing load. However, the size of
the component vectors has much less effect on processing load, so
the fact that the limit is not related to the size of the entities is also
explained. Thus, in terms of our relational model, there is a limit

on tensor rank entailed by the rapid growth of the number of ten- This is a binary relational instance and can be implemented by the
sor units as rank increases. Given that a short-term memory store tensor product
of capacity 4 is connected to a tensor-like system for processing,
the limit of 4 on store size is a consequence of the fact that for most vordered-fruit 3 vfirst 3 vapple 1 vordered-fruit 3 vsecond 3 vorange
cognitive tasks, processing of the objects in the store is a necessity. 1 . . . 1 vordered-fruit 3 vfourth 3 vpear
The links between storage and processing phenomena are Binding items into n-ary relations where n has a maximum value
worth exploring. In section 2, Cowan argues that the unity of con- 4. This can be implemented by a tensor up to Rank 5:
scious awareness implies the contents of attended channels should
be integrated or combined. Similarly, category clusters (discussed vfruit 3 vapple 3 vorange 3 vpear
in sects. 2.7 and 3.4.2) imply a link between instances of the cate-
gory. Cowan further contends, in section. 3.1.3, that the short term These representations have different characteristics. They per-
storage limit is observed only with items recalled in correct serial mit different retrieval operations, and impose different process-
positions. Given that the slots of relation are identified, serial po- ing loads. More important, at least some of these properties can
sition can be coded as a relation ordered-items (item 1, item 2, be captured by neural net models. The Rank n tensor would ex-
item 3, item 4 ). The observation of no limit with free recall would plain why processing load increases with the number of entities
then suggest that it is ability to represent the relation, rather than related, and consequently suggests why the capacity limit tends to
the items, that is subject to the limit. This would appear to be con- be low. However, the earlier representations are not sensitive to
sistent with the relational complexity theory of Halford et al. processing load in this way. It should be clear from these examples
(1998). Furthermore, it clearly points to explaining storage limits that storage and process are intimately related, and that a theory
in terms of complexity of relations that can be represented. This of capacity must include both aspects of computation. However,
would also explain the finding of Nairne (1991, referred to by while their interaction may be complex, it is not arbitrary. Our the-
Cowan in sect. 3.4.3) that errors occur up to three positions from ory specifies a unique set of properties for processes involving re-
the correct position. The reason would be that the items are rep- lations of different antics.
resented as a quaternary relation, which contains only four slots. Conclusion. Cowan has done the field a great service by show-
The further finding, in section 3.4.5 that participants could pre- ing that a broad range of observations is consistent with the limit
dict the seventh item from items 3, 4, 6 may also indicate that the of four entities that had been proposed previously by Halford et
task is represented as a quaternary relation. al. (BBS, 1998). However his claim to reduce processing capacity
These phenomena indicate links between entities that are im- to storage capacity is not substantiated. Furthermore he offers no
portant, but the nature of these links is not really clear, and the is- explanation for the limit, and glosses over the fact that at least one
sue is clouded by the lack of a well specified theory in Cowans pa- existing theory offers a potential explanation as to why the limit
per. Some properties of relational knowledge defined by Halford should be small.
et al. (1998) seem to be involved in the phenomena discussed
above, but it is not clear that they all are. We could define the re-
lational instances fruit (apple, banana, orange, pear) and fruit (ly-
chee, pineapple, passionfruit, guava), and so on. Organizing mem- Pure short-term memory capacity has
ory storage as quaternary relations in this way would account for
recall of items in clusters of four. However, it would also predict a
implications for understanding
lot of other properties of relational knowledge that Cowan has not individual differences in math skills
demonstrated. For example, relational knowledge has the prop-
erty of omni-directional access (Halford et al. 1998) which means Steven A. Hecht and Todd K. Shackelford
that, given any n21 components of a relational instance, the re- Division of Science, Department of Psychology, Florida Atlantic University,
Davie, FL 33314. {shecht; tshackelford}@fau.edu
maining component can be retrieved. Thus, given the quaternary
www.uni-bielefeld.de/ZIF/heiko.html
relation proportion (4, 2, ?, 3) we can determine that the missing
component must be 6 because it is necessary to complete the
Abstract: Future work is needed to establish that pure short-term mem-
proportion 4/2 5 6/3. However, it is far from clear that category ory is a coherent individual difference attribute that is separable from tra-
clusters share this property. If given a list [apple, banana, ? pear] ditional compound short-term memory measures. Psychometric support
there is no particular reason why we should recall orange. Thus for latent pure short-term memory capacity will provide an important
category clusters do not entail the kind of constraints that are en- starting point for future fine-grained analyses of the intrinsic factors that
tailed in relations. Another property of relational knowledge is influence individual differences in math skills.
that analogical mappings can be formed between corresponding
relational instances (Holyoak & Thagard 1995). Again, it is not Cowan presents a clear and convincing theoretical case for the
clear that analogies can be formed between category clusters. fixed capacity limit of three to five chunks in the focus of atten-
Storage is not a simple, unitary matter, but can take many forms. tion. Cowan has significantly advanced the field of memory re-
Furthermore, the form in which information is stored affects the search by providing a cogent analysis of the conditions that must
form in which it is processed. Some of the possibilities, together be met in order for memory storage capacity to be measured ac-
with possible implementation in neural nets, are: curately. Cowan presents an organized and impressive array of
1. Item storage implemented as a vector of activation values empirical data corroborating his theoretical claims. Cowans
over a set of neural units. claims rely on the idea that purer estimates of storage capacity
2. Associative links between items, implemented as connection in adults can be derived from existing sources of evidence. These
weights between units in different vectors. sources of evidence come from various methodologies, and con-
3. Superposition of items implemented as summation of item verge on the conclusion that a smaller than previously thought
vectors. This is tantamount to a prototype. chunk limit exists in the focus of attention.
4. Superimposed items bound to a category label, such as Cowans analysis provides researchers with a promising tool for
fruit(apple) 1 fruit(banana) 6 fruit(orange) 1 fruit(pear). This is relating capacity limits to individual differences in various kinds of
equivalent to a unary relation and can be represented by a Rank 2 human abilities; in this commentary, we address mathematical
tensor: thinking. Cowan distinguishes between memory measures that
vfruit 3 vapple 1 vfruit 3 vorange 1 vfruit 3 vorange 1 vfruit 3 vpear do, or do not, control for noncapacity-limited mechanisms. Com-
pound short-term memory tasks capture both memory storage
Item-position bindings: ordered-fruit (first, apple) 1 ordered- capacity and other sources of variance, such as strategic process-
fruit {(second, orange) 1, . . . , 1 ordered-fruit (fourth, pear)}. ing. An important limitation of traditional compound short-term

memory tasks is that the relative contributions of both pure short- pure short-term capacity size and performance on Logans (1988,
term memory and other noncapacity limited mechanisms to vari- experiment 4) alphabet arithmetic task. The alphabet arithmetic
ability in math skills are usually not readily determined. Only com- test is considered to be an analog measure of the acquisition of
pound memory tasks have been systematically investigated as simple arithmetic knowledge (Logan 1988). Cowan speculated
capturing memory processes that contribute to individual differ- that problems with addends of 14 can be visualized (i.e., held in
ences in math skills (see Geary 1993, for a review) the focus of attention) more clearly while problem solving, be-
There are at least three aspects of a complete account of indi- cause the addend sizes correspond to the number of chunks of in-
vidual differences in math skills (Hecht 1998). First, it is neces- formation that can be held in the focus of attention. In contrast,
sary to focus on specific subdomains of math skills (e.g., simple performance on problems with addends of five or more may be
arithmetic, fraction estimation), because different factors may in- hindered by pure short-term capacity limitations. Presumably, in-
fluence each subdomain of math ability. Second, the unique math- dividual differences in pure short-term memory capacity should
ematical knowledge (e.g., counting knowledge) needed to carry be associated with the effects of addend size on alphabet arith-
out specific problems in a subdomain should be determined. metic performance.
Third, the contributions of intrinsic factors such as memory ca- Obtained relations between pure short-term memory capacity
pacity on the efficiency with which mathematical knowledge is ac- and variability in math skills also may help disentangle the influ-
quired and carried out should be investigated. Another character- ences of biological and cultural factors on math attainment. Geary
istic of a complete account of variability in math skills is a (1995) makes a distinction between biologically primary abilities
description of the relative contributions of biological and envi- and biologically secondary abilities. Biologically primary abilities
ronmental mechanisms that influence the development of math are found cross-culturally and are designed by natural selection in
skills (Geary 1995). As Cowan alluded, pure short-term memory our evolutionary past. Biologically secondary abilities are not
capacity may be determined solely by biological factors optimized found in all cultures and require sustained formal training (e.g.,
by adaptive processes in human evolution. Quality of math in- reading, advanced calculus). It is likely that mean estimates of
struction would be an example of an environmental mechanism. pure short-term memory capacity size, and degree of individual
Cowans target article suggests important avenues of future re- differences in that construct, are uniformly found across cultures.
search that might lead to a complete account of variability in math In his description of teleological accounts of a pure short-term
skills. The first suggested line of research is the nature of individ- memory capacity limit, Cowan reviews evidence from several
ual differences in pure short-term memory capacity. The relations sources suggesting a plausible evolutionary function of a very lim-
between pure short-term memory capacity and variability in aca- ited capacity of chunks in the focus of attention. For example,
demic performance can be investigated only if pure estimates of Cowan summarizes evidence by Kareev (1995) that a limited pure
capacity can be measured as a distinct individual difference at- short-term memory capacity assists in the efficient detection of
tribute. Cowan notes that individual differences in capacity limits correlations between features in the physical world. A limited
appear to exist in seemingly disparate tasks such as Sperlings pure short-term memory capacity, shaped by evolutionary forces,
(1960) full report task, Cowan et al. (1999) unattended speech task, may currently be co-opted for many contemporary tasks such
and Luck and Vogels (1997) visual storage capacity task. Perfor- as biologically secondary mathematical problem solving skills.
mance on these seemingly disparate tasks should correlate if they Cowans analysis suggests a line of research for investigating po-
measure the same underlying pure short-term memory construct. tential indicators of co-optation in the domain of math skills. Co-
Confirmatory factor analysis (CPA) can be used to support em- optation of biologically primary memory ability would be sug-
pirically pure short-term memory capacity as a unique domain of gested by observed correlations between pure short-term memory
memory ability. Based on Cowans analysis, it is likely that pure capacity and biologically secondary math skills.
short-term memory and compound memory tasks will yield sepa- Cowans target article should stimulate important avenues of fu-
rate, though correlated, factors. The constructs should be corre- ture research toward demonstrating psychometric support for
lated to the extent that compound short-term memory task per- separate constructs of pure short-term memory and compound
formance is influenced by pure short-term memory capacity. An memory capacity. Current research focusing on individual differ-
important benefit of using latent variables is that correlations ences in mathematical thinking has much to gain from the kind of
among factors can be observed while controlling for sources of fine-grained analyses of memory capacity suggested by Cowan.
variance associated with task specific level and direction of effort The predictive validity of latent pure short-term memory capacity
or attention. CPA also can be used to determine whether separate would provide important progress toward understanding the bio-
versus central pure short-term memory capacity exists. For exam- logically primary factors that influence variability in math skills.
ple, CPA can be used to determine if visual versus verbal pure ca-
pacity tasks yield separate or singular constructs of memory stor-
age capacity.
If the construct validity in the psychometric sense of pure short-
term memory tasks is empirically established, then another line of
research could examine whether individual differences in the ca-
pacity limit of chunks in the focus of attention contributes to math
ability. Although relations between memory-related latent factors
and emerging variability in math outcomes have a been demon-
strated (see Hecht et al., in press), these predictors are nor as fine-
grained as the measures suggested by Cowan. That is, an impor-
tant limitation of extant research relating memory processes to
variability in math skills is that observed correlations do not indi-
cate which aspects of memory performance influence specific as-
pects of math ability (Geary 1993). Once separate latent variables
for pure short-term memory capacity and compound short-term
memory have been identified, the relative contributions of these
factors to individual differences in math skills can be assessed.
Thus, Cowans analysis suggests important starting points for
finegrained investigations of relations between intrinsic memory
abilities and variability in specific aspects of mathematical ability.
One place to start is suggested by Cowans speculations regarding

Dual oscillations as the physiological mated from the Sternberg slope, the upper limits of STM is about
basis for capacity limits 56 items. Future studies in which both theta and gamma are
measured simultaneously will allow a more precise estimate. Sev-
Ole Jensena and John E. Lismanb eral studies hint that this will be possible in the future. In a recent
aBrain Research Unit, Low Temperature Laboratory, Helsinki University of EEG study, Tallon-Baudry et al. (1999) have reported ongoing
Technology, Helsinki 02015 HUT, Finland; bDepartment of Biology, Volen gamma activity during the retention period of a delayed-match-
Center for Complex Systems, Brandeis University, Waltham, MA 02254. ing-to-sample task. The frequency of the gamma activity ranged
ojensen@neuro.hut.fi lisman@brandeis.edu from 24 to 60 Hz which is too broad to be used for estimating the
www.boojum.hut.fi/ojensen STM capacity. There have also been attempts to manipulate the
www.bio.brandeis.edu/lismanlab/faculty.html frequency of the gamma rhythm. By delivering periodic auditory
stimuli in the gamma range (or half the gamma frequency), Burle
Abstract: A physiological model for short-term memory (STM) based on and Bonnet (2000) sought to entrain the gamma generators in hu-
dual theta (510 Hz) and gamma (20 60 Hz) oscillation was proposed by mans performing the Sternberg task. Consistent with the oscilla-
Lisman and Idiart (1995). In this model a memory is represented by
groups of neurons that fire in the same gamma cycle. According to this
tory memory model they were able to increase or decrease the
model, capacity is determined by the number of gamma cycles that occur Sternberg slope by supposedly driving the gamma generators to
within the slower theta cycle. We will discuss here the implications of re- higher or lower frequencies.
cent reports on theta oscillations recorded in humans performing the In conclusion, it is now possible to envision experiments which
Sternberg task. Assuming that the oscillatory memory models are correct, will rigorously test oscillatory models of STM. It may be possible
these findings can help determine STM capacity. to determine by direct measurement how many gamma cycles oc-
cur during a theta cycle while a subject is actually performing a
In reading the target article by Cowan it is evident that it is prob- STM task. If this number correlates with storage capacity, as mea-
lematic to determine the STM capacity from psychophysical ex- sured psychophysically, it would lend credence to oscillatory mod-
periments alone. The main problem is to design experiments in els and provide insight into the absolute magnitude of capacity, in-
which the influence of chunking and long-term memory are con- dependent of assumptions about chunking.
trolled. The framework of oscillatory memory models (Jensen &
Lisman 1996; 1998; Lisman & Idiart 1995) may allow the capac-
ity of STM to be tested more directly. In these models multiple
representations are assumed to be kept active by a multiplexing
network where the dynamics are controlled by nested theta (5 10
The magic number four: Can it explain
Hz) and gamma (20 80 Hz) oscillations. A memory is represented Sternbergs serial memory scan data?
by a group of neurons that fire in the same gamma cycle. The set
of memory representations is sequentially reactivated, one repre- Jerwen Jou
sentation per gamma cycle, in each theta cycle. Hence, the num- Department of Psychology and Anthropology, University of Texas-Pan
American, Edinburg, TX 78539-2999. jjou@panam.edu
ber of gamma cycles per theta cycle determines the capacity of the
www.w3.panam.edu/~jjou
memory buffer.
The retrieval time from STM is measured by the Sternberg
Abstract: Cowans concept of a pure short-term memory (STM) capacity
method in which a set of S items is presented to a subject. After a limit is equivalent to that of memory subitizing. However, a robust phe-
few seconds retention the subject must press a button to indicate nomenon well known in the Sternberg paradigm, that is, the linear in-
whether a probe item matched one of the items on the list. The crease of RT as a function of memory set size is not consistent with this
models predict that the gamma period, Tgamma , determines the in- concept. Cowans STM capacity theory will remain incomplete until it can
crease in reaction per item (the Sternberg slope). account for this phenomenon.
In the initial model it was proposed that the Sternberg slope
equaled Tgamma (Lisman & Idiart 1995). In later work, which at- After almost half a century of being imbued in the doctrine of
tempted to account for the full distribution of the reaction time magic number 7, one is reminded by Cowans target article that
data, a correction term was introduced (Jensen & Lisman 1998). this number may not be as sure a thing as most people have be-
Since memory scanning cannot be initiated in the period of the lieved it to be. Cowans target article poses some serious chal-
theta cycle when the S items are activated a wait-time comes into lenges to a long established tenet about memory and provides the
play. The average wait-time is s! STgamma . Hence the corrected field with new ideas and the vigor necessary for the continued ad-
contribution to the retrieval time is s! STgamma 1 Tgamma S 5 vancement of our discipline. However, some well established facts
s# TgammaS resulting in the Sternberg slope: s# Tgamma . The Sternberg in the short-term-memory (STM) literature seem to be incom-
slope has been determined to 3540 msec/item, hence Tgamma < patible with Cowans characterization of the pure STM capacity
25msec ( fgamma 5 40 Hz) in multiple psychophysical studies. limit. The concept of pure capacity limit of 4 has to account for
While this strategy may make it possible to estimate Tgamma psy- these well known STM facts if it is to become a serious competi-
chophysically, there are not yet clear strategies for determining tor of, and a possible alternative to Millers (1956) theory of STM
the frequency of theta psychophysically. Thus, we cannot suggest capacity of 7 6 2.
how the framework of oscillatory models would provide an inde- Cowan argues that a pure measure of STM storage capacity,
pendent way of measuring storage capacity on the basis of purely uncontaminated by rehearsal (or the use of other mnemonic de-
psychophysical data. vices such as chunking) or the mental processing sometimes re-
The great advantage of the oscillatory memory models is that quired beyond simple retaining, can be obtained. The size of this
they can be tested by recording brain rhythms in humans per- capacity is 4. One crucial defining property of the information
forming STM tasks. So far there are no reports on ongoing gamma held in this pure STM storage capacity, according to Cowan, is its
activity measured during the Sternberg task. However, in several being the target of the momentary focus of attention. Any infor-
recent studies it has been possible to measure ongoing theta ac- mation within this window of attentional focus is fully activated
tivity in subjects performing the Sternberg task (Jensen & Tesche and fully accessible. In other words, Cowans concept of a pure
2000; Raghavachari et al. 1999). In these studies the theta fre- STM storage is a memory version of the perceptual process known
quency during the retention interval of the Sternberg task was as subitizing (Jensen et al. 1950; Jou & Aldridge 1999; Kaufman
measured at 7 8 Hz. The on- and offset of the theta activity cor- et al. 1949; Klahr 1973; Logie & Baddeley 1957; Mandler & Shebo
related with the events of the task, whereas the frequency was in- 1982), which is defined as a rapid, effortless, and yet highly accu-
dependent of the memory load. Applying the theta frequencies rate immediate apprehension of the numerosity of a small num-
measured in these experiments and the gamma frequency esti- ber (typically under five) of items. It is also in essence the same as

what Ebbinghaus called the window of simultaneous conscious- the memory representation and retrieval processes are automa-
ness (cited in Slemecka 1954). Cowan implies in many parts of the tized through overlearning).
target articles and indicates, with support from empirical data, Again, Cowans concept of a small STM window of direct and
that the reaction time (RT) for retrieving information from mem- simultaneous access of information must explain the above data to
ory sets within the size of the pure STM limit remains relatively be a viable theory. Specifically, it has to account for (1) the robust
low and constant, but shows a sudden large increase as the mem- linear increase of RT up to a MSS of 6 (a phenomenon suggesting
ory set size (MMS) exceeds 4, in parallel with the finding in per- a lack of simultaneous consciousness for the information) and (2)
ceptual quantification of a discontinuity in RT functions from the the leveling off of the RT functions as MSS exceeds 6 (which is the
subitizing range to the counting range (beyond 3 or 4 items) opposite of what Cowans theory would predict).
(Mandler & Shebo 1982). But, is there such a parallel? Also incompatible with the authors view was the serial position
The concept of a smaller-sized STM storage, which is the mo- confusion pattern, ironically cited (Nairne 1991; 1992) by Cowan
mentary focus of attention, or the window of simultaneous con- to support his concept of STM capacity. That is, the confusion oc-
sciousness, must account for a very robust phenomenon in STM curred mostly within 3 or 4 positions of the target item. This con-
literature if it is to become a viable theory of STM. The phenom- tradicts the basic concept of subitizing. If the information process-
enon is the linear increase of RT as a function of MMS, first ing within the small range of 4 items is like perceptual subitizing,
brought forth by Sternbergs seminal papers (Sternberg 1966; then no errors, or at most a minimum number of errors should oc-
1969) and later replicated by countless studies conducted in dif- cur, because perfect performance is a hallmark of subitizing.
ferent contexts and with different test materials (Sternberg 1975). Finally, Cowan attempts to define a condition under which a
If items reside within the attention-focused window of simultane- pure STM storage capacity can be measured, that is, one in which
ous consciousness, why does the time required to access or re- the memory system is neither overburdened nor under-taxed. This
trieve an item from this range increase linearly with MMS so con- seems to involve a very delicate balance between too much and too
sistently and reliably? Cowan presents some cases where the RT little resource demand. For the theory to be formalized, this deli-
for memory retrieval increases little until the MMS exceeds 4. cate balanced condition has to be more clearly spelled out. It is pos-
How can the concept of memory subitizing and the data cited by sible that a MMS of 4 is not the STM limit, but an optimal work-
the author in support of it be reconciled with the even larger body ing range of STM. Capacity is the largest possible amount of
of data that is consistent with the Sternbergs concept of serial information that STM can hold. If the focus of attention can be
memory scan? switched, as the author suggested, between two or three blocks of
Furthermore, contrary to the finding cited in the target article three items each (assuming the chunking is not formed by relying
of a shallow slope of RT functions associated with MMS of 4 or on long-term-memory knowledge, but by using temporal proxim-
smaller, and a steeper slope for MMS past 4, a considerable num- ity), and if all three chunks are available within a certain short pe-
ber of studies have shown the opposite pattern of results. In these riod of time (though not necessarily simultaneously as in the case
studies, the RT functions for MMS of 6 or smaller are typically of focused attention), why cant all three chunks be considered
characterized by steep linearity. But as the MMS exceeds 6, the parts of the STM? What is the a priori basis for limiting the STM
functions become essentially flat (Baddeley & Ecob 1973; Bur- capacity to the span of focused attention, or simultaneous con-
rows & Okada 1975; Okada & Burrows 1978) Those results were sciousness? Grouping several numbers together into a chunk by
interpreted as suggesting that a serial access mode was in opera- reading these numbers faster as a unit or delivering items at an
tion for small MSSs, but that a direct access mode is adopted when overall faster rate (as in Waugh & Norman 1965, cited in the target
MSS exceeds 6. Jou (1998) used a fixed set version of the Stern- article) increases the total amount of information that can be held
berg task with MMSs varying from 1 to 20 and memory-set items in STM by minimizing the loss of information over time. Why cant
randomly and repeatedly sampled from the 50 U.S. state names. it be justified to push the STM limit to its maximum by delivering
RT increased at about 66 msec/item up to the MMS of 6 and then the items at a faster rate? Cowan states that the presentation rate
1eve1ed off to an insignificant 7 msec/item past MSS 6. This was should not be too slow, either. Again, concerning the presentation
consistent with the findings of the above-noted studies. The at- rate, there seems to be a delicate balance in order to demonstrate
tention focus theme emphasized throughout Cowans target arti- a capacity of 4. What this exact rate of presentation is did not seem
cles would have predicted results of reversed pattern. to have been clearly specified in the target article either.
Jou and Aldridge (1999) had subjects estimate the serial posi-
tions of letters in the alphabet and the alphabetic distances be-
tween two letters. This task can be considered to involve a form of
memory quantification of some overlearned magnitude facts. The
results were that, for the serial position estimation, the RT/alpha-
betic serial position functions were linear and steep up to the se- Magical number 5 in a chimpanzee
rial position of 6 or 7, past which the RT functions turned essen-
tially flat. The RT functions for the alphabetic distance judgment Nobuyuki Kawai and Tetsuro Matsuzawa
showed basically the same pattern except that the steep linear por- Section of Language and Intelligence, Department of Behavior Brain
tion was reduced to a magnitude of about 4. Although these re- Science, Primate Research Institute, Kyoto University, Kanrin, Inuyama City,
Aichi Pref. Japan 484-8506. {nkawai; matsuzaw}@pri.kyoto-u.ac.jp
sults are highly counterintuitive in that memory quantification
www.pri.kyoto-u.ac.jp/koudou-shinkei/shikou/index.html
seems to operate in a serial fashion for small values, but in a par-
allel fashion for large values, they are consistent with Burrows and
Abstract: One of our recent studies has revealed that a numerically
Okada (1975), Okada and Burrows (1978), and Baddeley and trained chimpanzee can memorize a correct sequence of five numbers
Ecob (1973) findings. Jou and Aldridge (1999) concluded that shown on a monitor. Comparative investigations with humans show very
there is no memory subitizing, unlike in the perceptual domain. similar patterns of errors in the two species, suggesting humans and chim-
They suggested that there is a fundamental difference between panzee share homologous memory processes. Whether or not 5 is a pure
quantification in memory and that in perception because in per- capacity limit for the chimpanzee remains an empirical question.
ception, stimuli are physically present whereas in memory they
have to be internally represented, which consumes resources and Cowan has proposed a new magical number 4 in human adults
prevents subitizing from taking place. Perceptual subitizing, ac- from a careful reconsideration of short-term memory processes.
cording to Jou and Aldridge (1999), is a result of having an over- In his theoretical account of the capacity limit, Cowan suggests
abundance of attentional resources available. In memory, this that the capacity limit might have become optimized through
rarely occurs because of the resource demands made by the men- adaptive processes in evolution (sect. 4.1). If this is the case, we
tal representation process (the only exception perhaps being when can expect to find similar memory processes in non-human ani-

mals. Chimpanzees are good candidates for investigating the ori- memory systems may share homologous mechanisms (Kawai &
gins of our cognitive evolution, being the closest relatives of hu- Matsuzawa 2000a; in press).
mans among living creatures. Because masking occurred at the first touch, one might argue
Recently, we found that a numerically trained chimpanzee had that the memory span of the chimpanzee is four instead of five. As
a memory for numbers in several aspects similar to humans demonstrated above, both Ai and the humans had already planned
(Kawai & Matsuzawa 2000b). The chimpanzee called Ai, has more their response at the onset of each trial: there is thus the possibil-
than 20 years of experimental experience. Prior to the memory ity that the first number was included in the memorized sequence.
test, Ai learned to count dots on computer monitor or real ob- Even more, according to our recent test for six numbers Ais per-
jects and to select the corresponding Arabic numerals on a touch- formance was about 30%, significantly higher than the chance
sensitive monitor (Matsuzawa 1985). Ai also learned to order the level (0.8%). We do not deny the possibility that the pure capac-
numbers from zero to nine in sequence, regardless of the inter-in- ity limit of the chimpanzee might be less than four. Further ex-
teger distance (Biro & Matsuzawa 1999; Tomonaga & Matsuzawa perimental studies will be required determine this. Nevertheless,
2000). Utilizing her numerical skills, we set up a memory task. A because comparable data were obtained in Ai and the humans, our
set of numbers (e.g., 1, 3, 4, 6, and 9) was spatially distributed on study strongly suggests that if there are any pure capacity limit
a screen. Ai was required to touch the numbers in an ascending differences between the two species, they should be quantitative
order. Immediately after the selection of the lowest number (i.e., rather than qualitative. The essential point is that humans share
1), all the remaining numbers were masked by a white square. quite a similar memory process with chimpanzees.
Hence Ai had to memorize the numbers (now masked) accurately
to select the correct sequence. She reached more than 95% cor-
rect with four numbers and 65% with five, significantly above
chance in each case (17 and 4%, respectively). This indicates that
she could memorize the correct sequence of any five numbers
What forms the chunks in a subjects
(Kawai & Matsuzawa 2000b). performance? Lessons from the CHREST
The most interesting result concerned Ais response time. The computational model of learning
longest response times were obtained for the first number of the
sequence. Response times were shortest for the other numbers, Peter C.R. Lane, Fernand Gobet, and Peter C-H. Cheng
and did not differ from one number to another. Thus, her mean ESRC Centre for Research in Development, Instruction and Training,
reaction time of first response to a set size of five was 721 msec, School of Psychology, University of Nottingham, Nottingham NG7 2RD,
and then 446, 426, 466, and 41, respectively, for the remaining United Kingdom. {pcl; frg; pcc}@psychology.nottingham.ac.uk
www.psychology.nottingham.ac.uk/staff/{Peter.Lane; Fernand.Gobet;
four, (now masked) numbers. This pattern of responses is similar
Peter.Cheng}
to that of humans. For example, mean reaction times of five adult
humans in the test were 1,430, 524, 490, 741, 672 msec, for each
Abstract: Computational models of learning provide an alternative tech-
response, suggesting that both Ai and humans memorize the num- nique for identifying the number and type of chunks used by a subject in
bers and their locations before the first response (Biro & Mar- a specific task. Results from applying CHREST to chess expertise support
suzawa 1999; Kawai, in press). the theoretical framework of Cowan and a limit in visual short-term mem-
One may argue that both Ai and humans might use a rehearsal ory capacity of 3 4 looms. An application to learning from diagrams illus-
strategy during the longest reaction time preceding the first trates different identifiable forms of chunk.
choice. The accuracy of humans decreased when the numbers
were masked 750 msec after the initiation of the trial. However, Cowans theoretical framework (sect. 2) assumes that the focus of
Ais reaction times were almost half of those of humans, and they attention is capacity-limited, and that deliberately recalled [in-
remained approximately the same for the masked and unmasked formation] is restricted to this limit in the focus of attention. This
trials. Although rehearsal constitutes a major cause of compound framework is compatible with the EPAM/CHREST family of
STM estimate, rehearsal was impossible because Ais fast reaction computational models, and this commentary highlights the role
times seem incompatible with rehearsal. Thus, Ais performance that a model of learning can play in clarifying the nature of chunks.
in memorizing five items may reflect a pure capacity limit. CHREST (Chunk Hierarchy and REtrieval STructure) is a com-
There is a possibility that other mnemonic strategies were in- putational model of expert memory in chess players (Gobet 1998;
volved in the task. For instance, one might suspect that Ai used Gobet & Simon, in press), and is based on the earlier EPAM model
the configuration of the numbers as possible spatial cues for re- (Feigenbaum & Simon 1984) of perceptual memory. CHREST
sponding. The procedure ruled out this possibility however, be- possesses an input device (simulated eye), a visual short-term
cause locations were randomized across trials and all trials were memory (STM) for storing intermediate results (equivalent to
unique in each session, thereby demonstrating that long-term the focus of attention), and a long-term memory (LTM) based
memory did not contribute to the performance. In addition, a de- around a discrimination network for retrieving chunks of infor-
tailed analysis of error trials confirmed that neither Ai nor the hu- mation. Each chunk is learnt from information in the visual field,
mans used spatial cues. The majority of errors (84.1% for Ai, using the STM to compose information across one or more eye
84.5% for humans) consisted in skipping one number only (e.g., fixations.
selecting 1 36 instead of 1 34 69). The remaining errors The classic recall task (Chase & Simon 1973; Cowan, sect. 3.4.1;
were also independent of spatial factors. Most (87.5% for Ai, and De Groot 1946; 1978) has been used to show that subjects recall
82.0% for humans) consisted of selecting the highest number in information in chunks. The task requires the model/subject to ob-
the sequence (i.e., 1 9 or 1 39 instead of 1 34 69), regard- serve a display for a set time period, and then reconstruct the stim-
less of the spatial arrangement of the numbers on the screen. ulus from memory; in simulations, the chunks within the models
These trials were regarded as showing a recency effect because STM are used as the reconstructed response. In a study of chess
the highest numbers were the last to be processed in the pre- expertise, Gobet (1998) showed how the accuracy of the recon-
planned sequence. Even more interesting, the frequency of these structed position depends on the number and size of chunks
last numbers was proportional to the size of the greatest number: which the model identifies; the size of chunk depends on the level
more errors of this type were made when the sequence contained of expertise, but the number can be systematically varied, and a
a 9 as the last number than when it contained an 8, and that ten- value of 3 or 4 was found to best match the performance of differ-
dency remained for the lowest numbers. All these results suggest ent levels of player, providing further empirical support for the
that, like humans, Ai built up a linear representation of numbers, findings of Cowan. Also significant is that the better performance
from 0 to 9, and referred to it in performing the task. Similarity in of experts is explained by their use of larger chunks (typically, mas-
the error patterns for the two species moreover suggests that their ter chess players recall chunks of twice the size of average club

players), and the number and content of these chunks may be ex- that there is a 3-to 4-item limit on the simultaneous coding and re-
tracted from the model (see also Gobet & Simon 1998; in press) production of elements, at least in some domains. Cowan believes
Chase and Simon (1973) did, however, find that expert chess the same limit holds for sequentially displayed elements, namely,
players appeared to recall more chunks than novices. As discussed elements distributed over time rather than space. However,
in Gobet and Simon (1998), these findings do not contradict the Cowans estimates are based largely on indirect measurements.
existence of a fixed capacity limit, because additional factors affect Crucially, measures of retrieval speed from studies using sequen-
the subjects performance; in this case, the number of pieces which tial presentation provide direct evidence for a distinct representa-
the player can pick up. So, are the chunks observed in the subjects tional state associated with the focus of attention that is limited to
performance due to previously learnt information or to other fac- the most recently processed unit. These measures indicate, con-
tors relating to the task or cognitive performance? This question tra Cowan, that only one chunk is maintained across a dynamically
may be answered through a simulation of the learning process. The changing environment. One possibility is that the capacity of focal
role of learnt knowledge in producing chunks in performance is attention differs for simultaneously available elements arrayed in
currently being explored in a problem-solving version of CHREST space, and for representations encountered over time. If Cowans
(Lane et al. 2000a) which learns a diagrammatic representation analysis is correct, perhaps we can attend to more than one si-
for solving electric circuit problems. In Lane et al. (2000b) differ- multaneously presented element; however, we do not appear to be
ent computational models were analysed based on their respec- able to process more than one temporally extended event.
tive representational, learning, and retrieval strategies for hand- Cowans evidence. Cowan forwards, as an estimate of the ca-
ling high-level information. From these two studies, it is clear that pacity of focal attention, findings that the number of recalled
chunks observed in the models performance may arise from a items often converges on 34. However, recall performance is de-
number of causes. Three of the more apparent are as follows: termined by a confluence of factors other than the capacity of fo-
(1) A chunk may be observed in the output because of an ex- cal attention. Undoubtedly, these estimates partly reflect the re-
plicit representation in the systems LTM, which is the underlying call of representations outside focal attention, analogous to the
representation used in the EPAM/CHREST family of computa- way that serial position functions were classically argued to reflect
tional models. For example, Richman (1996) describe a chunk as output from both long-term and short-term. Further, recall is lim-
any unit of information that has been familiarised and has be- ited by forgetting that occurs over the learning phase and during
come meaningful. the recall process (e.g., Dosher & May1998). The number of items
(2) A chunk may be observed in the output because the input recalled, even when the preconditions enumerated by Cowan
has matched a stored chunk based on some similarity-based crite- (sect. 1.2) are met, provides at best an indirect estimate of the ca-
rion; this is familiar from neural network approaches. pacity of focal attention, and is equated with focal attention pri-
(3) A single chunk may be observed although it is based on a marily by assumption.
functional composition/decomposition of the stimulus and its sub- Retrieval speed. The claim that focal attention is distinct from
components. For example, subjects may retrieve and store multi- more passive memory representations implies that information in
ple chunks within their STM, but the performance based on these focal attention is accessed more immediately than information in
multiple chunks may then give the appearance of a single chunk. a passive state. Measures for access speed can provide direct evi-
The presence of three distinct processes yielding chunk-like be- dence for distinct representational states if access speed can be
haviour in such models clarifies how the observational character- measured for memories with different strengths (or probability of
istics of chunks inter-relate with learnt knowledge, and hence clar- access). Unfortunately, RT does not provide pure estimates of re-
ifies the connection between observed and learnt chunks. This trieval speed because it is affected by memory strength (e.g.,
connection assists in developing a deeper understanding of the ca- Dosher 1984; McElree & Dosher 1989; Wickelgren et al. 1980).
pacity limit, especially in areas where the subject is continuously However, retrieval speed can be directly measured with the re-
learning new chunks for composite objects. Most importantly, sponse-signal speed-accuracy tradeoff (SAT) procedure. In this
only by modelling the entire learning history of each subject can procedure, subjects are cued to respond at some time after the on-
we really attempt to probe the content and format of chunks ma- set of a test probe. With a suitable range of cue times, the full time
nipulated in STM, and thereby estimate STM capacity. course of retrieval is evaluated, providing measures of when in-
formation first becomes available, the rate at which information
accrues over retrieval time, and the asymptotic level of observed
performance. The asymptote reflects the probability of retrieval,
and provides an estimate of memory strength. When accuracy de-
The focus of attention across space parts from chance, the rate at which it grows to asymptote jointly
and across time measure retrieval speed. More accessible information should be
associated with an earlier intercept or faster rate, irrespective of
Brian McElreea and Barbara Anne Dosherb differences in asymptotic accuracy.
a
Department of Psychology, New York University, New York, NY 10003; Wickelgren et al. (1980) used a probe recognition task to ex-
b
Cognitive Science, University of California, Irvine, CA 92717; amine SAT time-course profiles for accessing representations in a
bdm@psych.nyu.edu bdosher@uci.edu list of 16 sequentially-presented items. Asymptotic accuracy de-
www.psych.nyu.edu/dept/faculty/mcelree/research.html creased monotonically with the decreasing recency of the tested
www.aris.ss.uci.edu/cogsci/personnel/dosher/dosher.html
item, indicating that memory strength systematically declines as
time or activity is interpolated between study and test. Crucially,
Abstract: Measures of retrieval speed for recently presented events show however, retrieval speed was constant across all serial positions save
a sharp dichotomy between representations in focal attention and repre-
sentations that are recently processed but no longer attended. When in-
the last, most recently studied position. Retrieval speed was 50%
formation is presented over time, retrieval measures show that focal at- faster when no items intervened between study and test. The most
tention and rapid privileged access is limited to the most recently recently studied item received privileged access. This finding has
processed unit or chunk, not the last 3 5 chunks that Cowan estimates been replicated with different procedures and materials, including
from various recall procedures. a Sternberg task (McElree & Dosher 1989), a forced-choice recog-
nition task (McElree & Dosher 1993), a paired-associate recogni-
Cowan presents a diverse array of evidence to support the claim tion task (Dosher 1981), and even when the task required judging
that the focal attention has a capacity of 3 4 chunks. Much of this whether a test item rhymed or was synonymous with a studied item
evidence comes from studies examining processing limits in multi- (McElree 1996). Related effects are found in judgments of re-
element displays in which all elements are simultaneously dis- cency (McElree & Dosher 1993) and the n-back task (McElree, in
played. These studies may provide good evidence for the claim press).

This sharp dichotomy in retrieval speed fast access for the
most recently processed item and a slower, but constant retrieval
speed for all less recently processed items provides direct evi-
dence for two representational states, one associated with focal at-
tention, and the other associated with memory representations
outside of the focus of attention. Although rapid access is usually
reserved for the most recently processed single item, McElree
(1998) found that rapid access accrues to multiple items if they
form a chunk or unit: the retrieval advantage extended to three
items if they were members of the most recently experienced cat-
egory. Further, the retrieval advantage is not bound to the last po-
sitions in a list, but instead reflects the last cognitive operation
(McElree 1997): an advantage is found for a non-recent category
when a category cue is used to retrieve and restore items to focal
attention.
Conclusions. Collectively, direct measures of retrieval speed
indicate that focal attention, associated with especially rapid re-
trieval, is more limited than the three or four items suggested by
the indirect analysis forwarded by Cowan. Measures of retrieval
indicate that we are only able to maintain one temporally extended
event or epoch in focal attention.
Figure 1 (McKone). Explicit short-term memory in continuous
old-new recognition (data from McKone, 1995, Experiment 4; av-
eraged across high frequency and low frequency word conditions).
Capacity limits in continuous old-new
recognition and in short-term vening items. This discontinuity between lags 2 and 3 corresponds
implicit memory to a short-term memory capacity of 4 items: at lag 2, the contents
of short-term memory are the current word, its first presentation,
Elinor McKone and two intervening items.
Division of Psychology, Australian National University, ACT 0200, Australia. Cowans exact prediction for the old-new recognition experi-
elinor.mckone@anu.edu.au www.psyc.anu.edu.au/staff/elinor.html
ment, however, was that the mean reaction time should be much
faster when the item had been presented within the most recent
Abstract: Using explicit memory measures, Cowan predicts a new cir-
3 or 4 items than when it was presented only earlier in the se-
cumstance in which the central capacity limit of 4 chunks should obtain.
Supporting results for such an experiment, using continuous old-new quence. This suggests that the discontinuity in RTs should take
recognition, are described. With implicit memory measures, Cowan as- the form a flat function up to 4 chunks (i.e., all items within the
sumes that short-term repetition priming reflects the central capacity focus of attention show equally fast RTs), followed by decay. The
limit. I argue that this phenomenon instead reflects limits within individ- data, in contrast, show smooth and rapid decay up to 4 items, fol-
ual perceptual processing modules. lowed by much slower decay. Thus, while Cowans prediction
seems to presume that all items in the focus of attention are
Cowan makes a prediction (sects. 1.2 and 4.3.7) regarding a new equally strong, the RT data instead suggest that older items fade
circumstance in which the capacity of short-term explicit memory from the limited capacity mechanism (at least without active re-
should be limited to 4 chunks. This is where subjects must only in- hearsal). This would seem to reconcile a capacity limit with the
dicate, as rapidly as possible, if a particular item had been included demonstrated importance of time-based factors when rehearsal is
in the stimulus set previously, and in which some items would be allowed (e.g., effects of articulation rate on word span tasks).
repeated in the set but other, novel items also would be intro- Turning to implicit, rather than explicit, memory tasks, Cowan
duced. Experiment 4 of McKone (1995) used such a design. Each cites my demonstration of short-term implicit memory (i.e., an ad-
trial presented a single word or pseudoword (e.g., mave), in conditional short-lived repetition priming advantage in tasks such as
tinuous sequences of 250 300 trials. Approximately 65% of trials lexical decision or word naming) as supporting evidence for a ca-
presented items new to the experiment; the other 35% were a re- pacity limit. However, it is not at all clear to me that short-term
peat of an item seen earlier in the list. Repeats occurred at vari- implicit memory is affected by the central capacity limit, rather
ous lags (i.e., number of intervening items: range 0 23). The than by module-specific processing factors. Theoretically, each
task was old-new recognition to every trial in the sequence, with item in turn in a lexical decision or naming task must fall within
reaction time for correct old responses as the dependent measure. the focus of attention to allow a response to be made to that item.
This experiment appears to satisfy Cowans general criteria for Unlike explicit memory tasks, however, no overt comparison of
producing a pure, rather than compound, capacity measure: the more than one item is ever required, and so it is not necessary that
fact that half the items were nonsense words and the lack of par- both presentations of the target are simultaneously held in Con-
ticular semantic associations between successive words, should scious awareness. Using Cowans terminology (sect. 4.3.6), an im-
limit chunking across items; the presentation rate (2 secs per trial) plicit memory task requires only chunks (e.g., x,y,z,x), and not
and the very long lists should limit active rehearsal. chunks in relation to some concept (e.g., recently seen [x,y,z,x]),
Results (see Fig. 1) were both consistent and inconsistent, with as does explicit memory.
Cowans predictions. A basic capacity limit of 4 items was sup- Empirically, McKone (1995) dissociated the apparent capac-
ported. In calculating the capacity, my logic is that (1) to be com- ity of short-term repetition priming from that of short-term ex-
pared with previous items, the current stimulus must be in the fo- plicit memory. With written stimuli, the explicit recognition mem-
cus of attention, (2) any previous item still in the focus of attention ory experiment described above obtained the 4 chunk capacity
should be recognised more quickly as old than an item which has limit for both words and pseudowords; however, implicit mem-
been pushed out, and (3) some sort of discontinuity in RTs should ory in lexical decision and naming, produced a purported capac-
therefore appear at the point at which the capacity limit is ity limit of 5 items for words but only 2 items for pseudowords.
reached. Figure 1 indicates that repeats at 0, 1, and 2 intervening McKone and Dennis (2000) then demonstrated that implicit ca-
items produce noticeably faster RTs than repeats at 39 inter- pacity varies consistently with the demands of visual versus audi-

tory word identification. Unlike (type) written words, the ambi- must be delivered, presumably via learned connections, to widely
guity of a continuous speech signal is so severe that a word can of- scattered neurons. In other words, if attention is to be directed to
ten be understood only following the arrival of several successive a book, the attention signal must find the diffuse cloud of neurons
speech segments. Thus, the auditory word identification system that represent visual features common to books.
might need to retain more items simultaneously active than does It is, thus, clear that the neural substrate of attention is not a
the visual word identification system and, since real words often unitary system that can be pointed like a spotlight or a camera,
initially sound like nonsense words, pseudowords might need to much less a static process into which peripatetic images can be di-
be retained as long as real words. Consistent with these proposals, rected. Apparently it must be a highly organized system of cen-
short-term repetition effects indicated capacities of 4 for words trifugal paths, every bit as specific as the centripetal sensory paths
and 2 for pseudowords with written presentation, but at least 5 that it modulates (Milner 1999). At least some of the neural activ-
6 (and possibly more) for both words and pseudowords with spo- ity corresponding to attention originates in the response planning
ken presentation. It is not obvious why a single central capacity mechanism where the intentions of the subject are elaborated.
should be responsible for these very different limits. Neural representations of objects, related either innately or
Overall, Cowan has provided convincing evidence of a central through prior experience to the task in hand, need to be selected
capacity limit (of 4 chunks) in short-term explicit memory, pre- by specific facilitation as part of the planned response.
sumably corresponding to the focus of attention. The data he re- According to the above account, what Cowan calls the focus of
views, however, do not directly address the question of whether or attention must vary widely depending on motivation, or the task.
not individual processing modules (perceptual, cognitive or mo- Thirst, for example, facilitates the sensory input from all sources
tor) also have their own capacity limits outside the focus of at- of water known to the subject, increasing the probability that if
tention. Clear theoretical reasons exist for limits on the number of one of them happens to be present it will gain control of the re-
items thatcan be simultaneously active within a single domain sponse mechanism. In the absence of a matching sensory input,
since left over activation from previous items will interfere with the dominant thirst-quenching intention facilitates items in long-
identification of the currently-presented item: in visual word term memory with which it has acquired associations. The more
recognition, many models in fact assume a capacity limit of only objects that have been associated with slaking thirst, the more
one item (the current stimulus kills all traces of previous words). widespread the initial attentional facilitation. Any input amplified
by attentional facilitation is then likely to determine the response
that is released.
In most memory-span measurements, the subjects task is to re-
call recently heard or read words. These must be distinguished
Magical attention from all the other words in the subjects vocabulary by short-term
Peter M. Milner changes in the thresholds or connections of corresponding neu-
rons. At a given moment during recall only one word is being re-
Department of Psychology, McGill University, Montreal, QC, H3A 1B1,
Canada. pmilne@po-box.mcgill.ca
leased to the motor system for pronunciation (or writing). One
problem is to discover why it is not the most recently heard word
Abstract: Cowan postulates that the capacity of short-term memory is lim-
(which presumably retains the strongest trace), another is to dis-
ited to the number of items to which attention can be simultaneously di- cover why the system can cope with no more than about 4 or 5
rected. Unfortunately, he endows attention with unexplained properties, words. Are the later words of a series less effectively tagged at the
such as being able to locate the most recent inputs to short-term memory, time of storage or does recall of the earlier words interfere with
so his theory does little more than restate the data. retrieval of the later ones? How important a factor is decay of the
trace with time? Can rehearsal of the earlier words ever be com-
Cowans thorough review of memory-span data provides convinc- pletely prevented? It seems to me that the limit of four words is
ing evidence that in the absence of rehearsal the span is about 4 at least as much to do with decay and interference of the short-
items. This information is useful, but how much does it further our term synaptic changes than anything related to attention.
knowledge to be told that the span is limited because the focus Another datum cited by Cowan in support of his theory is that
of attention can hold only about 4 items at a time? the subitizing limit is about 4. He bypasses the difficult problem
Cowan adopts Lisman and Idiarts (1995) theory that attention of how items simultaneously present within his postulated span of
is dependent on oscillations of cortical potentials. A cortical wave attention are summed to elicit a numerical response. Of course he
of about 10 Hz is supposed to select items from an unlimited is not alone in not having solved that problem, but it is at least pos-
short-term store; other wavelets at a frequency of about 40 Hz sible that the process involves discriminating a signal generated by
then select one item each. Assuming that such a process were neu- x objects from that generated by x 6 1 objects. The ratio of x to x
rally feasible, it still does not explain why successive waves select 6 1 may become too small to permit discrimination when x is
the same items, moreover the wave frequencies vary considerably, greater than about 4. I am confident that most people are able to
as Cowan points out. Apart from his uncritical recourse to this the- discriminate instantly between 10 objects and 20 objects without
ory of attention to explain the magical figure of 4, Cowan confines eye movements or counting, which might be interpreted as indi-
his discussion to cognitive models, so it may be unfair to criticise cating that at least 20 objects can simultaneously lie within
the theory on neural grounds. Nevertheless, I believe that con- Cowans focus of attention in some circumstances.
sideration of the constraints imposed by neurophysiology may be
helpful.
In the first place, it is unlikely that sensory information can be
moved to a focus of attention except in a metaphorical sense. At-
tention may change the intensity with which visual neurons re-
spond to a stimulus (Fuster et al. 1985; Moran & Desimone 1985),
but not their location. Furthermore, neurons fired by objects in
different parts of the visual field are spatially segregated only in
the primary visual cortex. At higher levels of the visual system,
neurons have large overlapping fields (Gross et al. 1974; Miyashita
1993). The relative positions in space of the objects they represent
do not determine the cortical location of the neurons. Thus indi-
vidual items in the visual field cannot be selected by spatially di-
rected attention. At this level of the visual path, selective attention

Nothing left in store . . . but how do we guistic estimates of the articulatory loop suggest that its capacity
measure attentional capacity? is larger in Chinese than in English (Cheung & Kemper 1993),
and in turn, larger in English than in Italian (Morra et al. 1993)
Sergio Morra this seems rather paradoxical, for a supposedly universal compo-
Universit di Genova, DISA-Sezione Psicologia, 16126 Genova, Italy. nent of the minds architecture.
morra@nous.unige.it Explanatory alternatives to the articulatory loop have been sug-
gested for the word-length effect, such as output interference,
Abstract: I compare the concepts of activation and storage as founda- proactive interference, and complexity of speech programming
tions of short-term memory, and suggest that an attention-based view of (see Brown & Hulme 1995; Caplan Rochon & Waters 1992;
STM does not need to posit specialized short-term stores. In particular, no Cowan et al. 1992; Henry 1991; Nairne et al. 1997; Service 1998).
compelling evidence supports the hypothesis of time-limited stores. Iden- In addition, various studies have found that verbal STM span is af-
tifying sources of activation, examining the role of activated procedural fected by variables that have little effect on articulation rate, such
knowledge, and studying working memory development are central issues as word familiarity, frequency, grammar class, semantic variables,
in modelling capacity-limited focal attention.
and order of stimulus words. Hence, the articulatory loop model
has either been dismissed, or transformed (e.g., by Burgess &
Cowans main thesis is that short-term memory depends on at- Hitch 1997) into something radically different from the original
tention. The focus of attention is capacity-limited . . . Any infor- time-limited store. Furthermore, estimates of loop capacity are af-
mation that is deliberately recalled, whether from a recent stimu- fected by use of span procedure versus supra-span lists (Morra
lus or from long-term memory, is restricted to this limit in the 1990; Mona & Stoffet 1991; Nicolson & Fawcett 1991). Thus, the
focus of attention . . . The same general principles of activation only supposedly precise estimate of a supplementary storage mech-
and de-activation might apply across all types of code (sect. 1). So anism may have been an artifact, obtained fortuitously from supra-
far, so good. span lists of English words.
However, Cowan often calls store the capacity-limited focus At this point, I think, we can abandon the idea of specialized
of attention (whereas Cowan 1988, made it clear that it is an acti- short-term stores, and retain instead the view of working memory
vated portion of long-term memory), and leaves as an open issue as the activated part of LTM. Apparently separate STM modules
the existence of supplementary storage mechanisms, which . . . can be regarded as epi-phenomena of LTM modularity. Instead of
are time-limited rather than capacity-limited (sect. 1). I think that searching for different supplementary stores, we can think of dif-
an attention-based view of STM entails that it is regarded as an ac- ferent sources of activation; that is, working memory must be seen
tivated part of LTM, which is in turn incompatible with the con- as broader than the capacity-limited focus of attention, because
cept of short-term storage. (Connecting an activation-based cen- LTM units may also be activated by other sources. These may in-
tral executive with specialized slave stores has been a long-standing clude current perceptual input, associative learning, top-down
problem for Baddeleys [1986] theory.) processes from higher-order cognitive structures (e.g., Case 1974;
Storage and activation are obviously metaphors, but not so Pascual-Leone 1987), and of course, residual activation of items
innocent and vague as to be interchangeable. They have empiri- that have recently been activated by the capacity-limited atten-
cally distinguishable consequences. tional mechanism.
(1) If there are short-term stores, performance should never be If we construe in this manner working memory and its relation
improved by occupying them with irrelevant materials (unless this to the capacity-limited focal attention, then we can ask further
prevents use of inadequate strategies: Brandimonte et al. 1992). questions. One is whether this view accounts for short-term mem-
In an activation-based system, instead, an irrelevant memory load ory phenomena traditionally explained in terms of storage. It seems
may sometimes pre-activate structures and thus enhance perfor- so; for instance, Morra (2000) has presented a neo-Piagetian
mance. The latter proved to be the case (e.g., Hellige & Cox 1976; model of verbal STM that does not include any time-limited spe-
van Strien & Bouma 1990). cialized store.
(2) Storage models must specify flow of information transfer, A second question concerns the content of working memory.
whereas activation models must specify time course of activation. Cowan implicitly suggests that it only includes declarative knowl-
Numerous priming paradigms lend themselves to an activation ac- edge, as one can infer from his mention of parietal lobes. How-
count. Detailed models have been proposed for the time course ever, LTM includes procedural as well as declarative knowledge,
of activation in both positive and negative priming (e.g., Houghton and one may assume that procedural knowledge also needs to be
& Tipper 1994; Neely 1991). I am not aware of equally powerful activated by attentional resources. In a neuropsychological per-
models of priming as information transfer among stores. As An- spective, Moscovitch and Umilt (1990) conceived working mem-
derson (1983, p. 21) put it: The results on associative priming ory as whatever processes are currently active (without distin-
have shown us that the amount of information brought into work- guishing between anterior or posterior parts of the brain) and
ing memory, at least temporarily, is very large. suggested that its limits are set by the resources necessary for
(3) If supplementary stores exist, then one should specify ei- mantaining information and operating on it. A corollary is that, if
ther their capacity or duration. Numerous short-term stores had also procedural information is considered (as some neo-Piagetian
been proposed in the literature, but only for one was a clear ca- theoreticians suggested; e.g., Pascual-Leone & Johnson 1991),
pacity estimation made. This was the articulatory loop, deemed then the estimated size of focal attention may be more than 4
to hold as much phonological material as can be uttered in about chunks.
2 seconds (Baddeley et al. 1975; for converging evidence see Bad- A third question concerns individual and developmental differ-
deley 1986; Hulme & Tordoff 1989; Schweickert & Boruff 1986). ences and their measurement. An answer to measurement prob-
Because this seems to be the only advantage gained by the stor- lems partly depends also on the assumptions one makes regarding
age view, it is worthwhile to examine it more closely. whether only declarative or also procedural information is
Early claims of falsification of a time-limited articulatory store counted. These questions are central to neo-Piagetian theories.
(Morra 1989, 1990; Morra & Stoffel 1991) remained unpublished. Different positions have been expressed in the debate on capac-
Journal reviewers discarded them, sometimes even on the grounds ity measurement, for instance on whether an average person at the
that the results were not credible, although some authors (e.g., highest point of cognitive development has a capacity of 4 or 7
Anderson & Matessa 1997) trusted them well and quoted them units (Case 1985; 1995; Ha1ford 1993; Morra 1994; Morra et al.,
extensively. However, evidence has continued to accumulate indi- in press; Pascua1-Leone 1970; Pascual-Leone & Baillargeon
cating that, contrary to Baddeley et al.s (1975) prediction, the ra- 1994). Clearly, the last word in this debate has not yet been spo-
tio of verbal recall to articulation rate is not a constant (e.g., Henry ken. Some results of my own research (Morra et al. 1988; 1991)
1994; Hulme et al. 1991; Morra 2000). The few available cross-lin- suggest, however, that early adolescents have a capacity of 5 or 6

units. This sort of developmental results may suggest that an av- ticular value of x, derived by letting L 5 (1/m) in Equation 2, will
erage adults capacity possibly spans over 6 or 7 chunks of infor- be labeled A, standing for actual maximum memory span.
mation. An example will illustrate these numerical values. If the target
material consisted of digit trigrams (for example, 2 1 8 or 6 3 9),
in which each individual digit has been drawn (randomly, with re-
placement) from the population of the ten digits 0 to 9, then the
probability, m, that a target trigram will be immediately followed
Partial matching theory by a target trigram bearing the same first digit will be .10 (one-
and the memory span tenth). Murray et al. (1999) argued that the memory encoding of
digit trigrams by participants in this task was indeed often in terms
David J. Murray of the first digits of those trigrams, especially if presentation were
Department of Psychology, Queens University, Kingston, Ontario K7L 3N6, purely visual with no auditory components. According to Equa-
Canada. murrayd@psyc.queensu.ca
tion 2, even if a list were extremely long, the theoretical maximum
number of (non-overwritten) memory representations of those
Abstract: Partial matching theory, which maintains that some memory
targets at the time of the probe would be approximately (1/m),
representations of target items in immediate memory are overwritten by
others, can predict both a theoretical and an actual maximum memory that is, approximately ten. But the actual number of memory rep-
span provided no chunking takes place during presentation. The latter is resentations available to the participant at this time, as determined
around 4 6 2 items, the exact number being determined by the degree of by constraints on conscious experience that are not yet under-
similarity between the memory representations of two immediately suc- stood, would be obtained by setting L 5 (1/m) 5 10 in Equation
cessive target items. 2, yielding an x-value of 6.5132. That is, the value of A, the actual
maximum memory span, would be 6.5132.
Cowans wonderful target article suggests that there is a pure Figure 1 shows predicted values of the theoretical maximum
limit of about four items in immediate memory. He notes that we memory span (1/m) and the corresponding actual maximum
do not, as yet, have an explanation for this limit. In this commen- memory span (A) for values of m ranging from .1 to .5. The range
tary, I shall suggest that an explanation can be derived from par- of in-values from m 5 .1 to m 5 .3 has been boxed off to show
tial matching theory as described by Murray et al. (1998). that, in this range, the predicted A-values lie between approxi-
This theory was developed with the intention of predicting hit mately 2 and 6, that is, they lie in the range 4 6 2.
and false alarm rates in immediate probed recognition tasks of the For single randomly selected digits, it was shown above that A
type investigated by Wickelgren and Norman (1966) and Stern- 5 6.5132. But other measures of memory span for single digits
berg (1966), among others. In this task, a sequence of L target usually provide estimates of the memory span for digits that ex-
items is presented, followed by a probe item that is either old or ceed 6.5132; for example, Cavanagh (1972), on the basis of a meta-
new with respect to the target list. The participants task is to judge analysis of previous reports, gave the traditional memory span for
whether the probe is indeed old or new. digits as 7.7. But, as Cowan has documented in impressive detail,
In partial matching theory, m is defined as the probability that most participants, upon hearing or seeing a sequential list of dig-
two adjacent target items share a predetermined feature of im- its, will bring to bear, on the process of the memorizing of that list,
portance in memory encoding. Following Neath and Nairne various techniques of chunking and associating that will expand
(1995), it was assumed by Murray et al. (1998) that the presence the number correctly recalled in order, after one presentation,
of this common feature would entail that the second target would from 4 6 2 to 7 6 2 or even more.
overwrite the memory representation of the first target. Murray Cavanaghs estimate that the traditional memory span for digits
et al. defined x to be the number of memory representations of is 7.7 is therefore almost certainly based on data that were not free
targets available at the time of onset of the probe and predicted from having been grouped. The actual maximum memory span for
that:
x 5 S (1 2 m)i for i 5 0 to i 5 (L 2 1) (1)
Equation 1 can be written alliteratively as

x 5 (1/m) [1 2 (1 2 m)L] (2)
As L tends to infinity, it can be seen from Equation 2 that x tends

to (1/m). According to partial matching theory therefore, the
maximum possible number of memory representations available,
at the time of the probe, of a very long list would tend towards
(1/m).
Experimental evidence that the theory can predict hit and false
alarm rates in immediate probed recognition tasks was provided
by Murray et al. (1989; 1999). Preliminary evidence that the the-
ory might also predict recognition latencies in this task was pro-
vided by Boudewijnse et al. (1999) in the course of their exposi-
tion of Herbarts (1824/1890) theory of how mathematics can be
applied to the prediction of how Vorstellungen (ideas) enter and
leave consciousness. Figure 1 (Murray). The theoretical and actual maximum mem-
However, the terms (1/m) and x can also be considered to be ory spans as predicted by partial matching theory. The variable-
measures of accuracy in immediate memory tasks generally. The name x on the ordinate denotes the number of non-overwritten
term (1/m) could represent a theoretical maximum memory memory representations at the time of the probe. The variable-
span, given a memory system subject to overwriting. If L itself name m on the abscissa denotes the probability that the memory
were set to be (1/m) the corresponding value of x would represent representation of a target item T shares a common feature (pre-
an actual maximum memory span, namely, the number of target determined by the experimenter) with the memory representa-
items out of (1/m) target items that had left memory representation of the target item immediately following. The value of the
tions that had not been overwritten, and were therefore, still ac- variable named A is given by (1/m)[1 2 (1 2 m)(12m)], as ex-
cessible to consciousness when an old probe appeared. This par- plained in the text.

digits of 6.5132 estimated by partial matching theory is a measure mains controversial (Cunningham et al. 1993; Healy & Cunning-
that presumes a minimum of chunking as the participant hears or ham 1995; Muter 1995).
views the target list under investigation. In practice, participants If the experiments of Peterson and Peterson and others like
usually do report more than A target items because the partici- them did indeed tap more than STM, then many questions remain
pants have succeeded in grouping or chunking the target items at unanswered regarding the nature of forgetting from STM, and are
the time of their presentation. To the references on grouping and not covered in Cowans target article. What is the typical rate of
chunking given in Cowans target article, we can add contributions forgetting? (This will undoubtedly depend on various circum-
by Slak (1970), Thompson et al. (1993), and Murray (1995, pp. 92 stances, just as the capacity does, but it may tend to be at a certain
105). level, just as capacity tends to be a certain chunk-size.) What is
the shape of the forgetting curve, and how does it compare to the
ACKNOWLEDGMENT shape in long-term memory (Rubin & Wenzel 1996)? Does the
Preparation of this manuscript was supported by Social Science and Hu- forgetting curve depend on the nature of the information re-
manities Research Council of Canada grant 410 99 0831. The author is membered (e.g., Murdock & Hockley 1989)? What are the roles
most grateful to Andrea Kilgour and Louise Wasylkiw for research assis- of decay, displacement, and interference (Laming 1992)? What is
tance, and to R. Kalin, Chair of the Department of Psychology at Queens the role of inter-chunk similarity (Posner & Konick 1966)? How
University at the time this research was initiated, for providing the neces- important are the expectations and needs of the rememberer (An-
sary computer facilities. derson et al. 1997)? What is the effect of the nature of the dis-
tracting activity (e.g., verbal vs. nonverbal, level of difficulty)? Can
forgetting from STM be instant, if an extremely salient multi-
modal event occurs? Almost all of the research on such questions
has been performed under conditions in which secondary mem-
The nature of forgetting ory contamination was likely, because the participants expected to
be tested after a filled retention interval.
from short-term memory Cowans paper usefully elucidates many issues regarding STM.
Paul Muter However, a comprehensive account of STM should surely include
treatment of the nature of forgetting after attention has been di-
Department of Psychology, University of Toronto, Toronto, Ontario M5S 3G3,
Canada. muter@psych.utoronto.ca
verted.
www.psych.utoronto.ca/~muter
ACKNOWLEDGMENTS
I thank Bennett B. Murdock, Jr. and Jay W. Pratt for helpful comments.
Abstract: Memory and forgetting are inextricably intertwined. Any ac-
count of short-term memory (STM) should address the following ques-
tion: If three, four, or five chunks are being held in STM, what happens
after attention is diverted?
Assuming that the central thesis of Cowans target article is cor- Long-term memory span
rect, the question remains: Given that three, four, or five chunks
are in STM, are negligibly registered in long-term episodic mem- James S. Nairne and Ian Neath
ory, and have been erased from sensory memory, what will be the Department of Psychology, Purdue University, West Lafayette, IN 47907-
1364. {nairne; neath}@psych.purdue.edu
nature of the forgetting when attention is diverted? In the third
www.psych.purdue.edu/{~nairne; ~neath}
paragraph of his paper, Cowan raises the issue of the nature of for-
getting from STM, and somewhat peremptorily dismisses it as
Abstract: Cowan assumes that chunk-based capacity limits are synony-
nearly intractable (para. 3), and beset with difficulties, such as mous with the essence of a specialized STM mechanism. In a single ex-
the apparent unresolvability of the decay issue (para. 3). Is the periment, we measured the capacity, or span, of long-term memory and
question of the nature of forgetting from STM any more in- found that it, too, corresponds roughly to the magical number 4. The re-
tractable than the question of the capacity of STM? sults imply that a chunk-based capacity limit is not a signature character-
Of course, some issues regarding the nature of forgetting from istic of remembering over the short-term.
STM are covered, both explicitly and implicitly, in Cowans paper,
but the above question is largely ignored, and is a remarkable la- Long-term memory span. As advocates of unitary approaches
cuna in the discussion. Memory and forgetting are always inextri- to memory, we applaud Cowans efforts to identify general
cably intertwined. mnemonic principles. His heroic review of the literature has pro-
Over the decades there have been hundreds of attempts to an- duced what appear to be remarkably consistent short-term mem-
swer approximations to the above question. Many of these at- ory capacity estimates. Although it would easy to quibble with the
tempts have been concerned with the rate of forgetting from selective nature of his review for example, Tehan and Hum-
STM. A study by Peterson and Peterson (1959) was quite typical: phreys (1996) report data counter to the claim that one can ob-
As Cowan mentions in passing, in a serial recall task Peterson and serve proactive interference only if there are more than four items
Peterson found severe forgetting of three letters after 18 seconds in a list we have chosen to focus our limited attention here on
of distracting activity. This study is often cited as indicating the more general issues.
duration of short-term memory (e.g., Solso 1995). Muter (1980), Cowan defines memory storage capacity operationally as the
however, argued that in the Peterson and Peterson experiments maximum number of chunks that can be recalled in a given situa-
and experiments like them, participants were relying on more than tion. Given this definition, a few concerns arise. First, almost all
STM, because they knew that they were going to be asked to re- the cited studies measure the number of items that can be re-
call the to-be-remembered items after an interval filled with dis- trieved rather than the number of items that can be stored (sect.
tracting activity. Theory (e.g., Craik & Lockhart 1972) and data 1. 1). As a result, bottlenecks in the retrieval process could well
(e.g., Jacoby & Bartz 1972; Watkins & Watkins 1974) suggest that lead to an underestimation of true storage capacity. Second, in vir-
if participants know they are going to be tested after a retention tually every case Cowan examines, the tasks require some form of
interval filled with distracting activity, secondary memory traces order or relational processing. For example, in the prototypical
are likely to be formed. When subjects expect to be tested after a case of memory span it is necessary to remember both the pre-
filled retention interval rarely or never, there is evidence that se- sented items as well as their ordinal positions in the list. It is un-
vere forgetting occurs after approximately 2 seconds, (Marsh et al. clear, as a consequence, whether the capacity limits apply to item
1997; Muter 1980; Sebrechts et al. 1989), though this finding re- information, order information, or to some combination of both.

Third, and most relevant to the remainder of our discussion,

Cowan assumes that chunk-based capacity limits are synonymous
with the essence of a specialized STM mechanism (sect. 1.1). Al-
though it is certainly possible that STM has such a limit, how can
we be certain that the same sort of capacity limits are not charac-
teristic of other memory systems? Does the magical number 4 ap-
ply only to the specialized STM mechanism, or might it apply as
well to other memory systems, such as long-term memory? To the
majority of memory theorists, Cowan included, the capacity of
long-term memory is assumed to be essentially unlimited. How-
ever, to our knowledge long-term memory span has never been
measured precisely, at least using the procedures and inclusion
criteria adopted by Cowan. What then is the storage capacity of
long-term memory?
In the typical memory span experiment, capacity is estimated
by requiring subjects to recall lists of various lengths and then pin-
pointing the list length that produces correct performance on 50%
of the trials. In principle, there is no reason why this procedure
cannot be adapted to a long-term memory environment. So, we
presented lists of various lengths to subjects and tested their abil-
ity to remember the lists 5 minutes later. To obtain a valid estimate
of storage capacity, Cowan argues, chunk size needs to be con-
trolled. Mnemonic strategies, such as rehearsal, can lead to hid-
den higher-order chunking that produces an overinflated esti- Figure 1 (Nairne & Neath). Long-term memory span: the pro-
mate of capacity. To prevent such strategies, we used unrelated portion of words correctly recalled for long-term memory as a
word lists under incidental learning conditions. function of list length.
Method.
Subjects. Two hundred and five Purdue University undergrad-
uates volunteered to participate in exchange for credit in intro- memory system. Instead, span limitations (the magical number 4)
ductory psychology courses. might simply be a characteristic of information load there is only
Materials and design. The stimuli were 22 concrete nouns from so much information that we can order correctly regardless of the
Paivio et al. (1969) with approximately equal length and ratings of time scales involved.
concreteness, imageability, and frequency. The subjects were as-
signed to one of two groups: Group 1 (N 5 105) received list
lengths of 2, 4, 7, and 9, whereas Group 2 (N 5 100) received list
lengths of 3, 5, 6, and 8.
Procedure. Subjects were tested in small groups and were in- Where the magic breaks down: Boundaries
formed that we were interested in obtaining pleasantness ratings and the focus-of-attention in schizophrenia
for several lists of items. Each word was pronounced out loud by
the experimenter at a rate of 1 word every 3 seconds, and each list Robert D. Oades and Boutheina Jemel
was separated by 5 seconds at the end of which the experimenter Biopsychology Group,University Clinic of Child and Adolescent Psychiatry,
said Next 1ist. Order of list length was partially counterbalanced 45147 Essen, Germany. oades@uni-essen-de; bjemel@excite.com
and the order of the words was random and different for each www.biopsychology.uni-essen.de
group. After rating the words, a geometric filler task was per-
formed for 5 minutes. Following this, subjects were given a sheet Abstract: The boundaries, the influences on, and consequences of a short-
term memory (STM) capacity of 4 leads us to consider global versus local
with all the words in their appropriate lists, with the words in al-
processing. We argue that in schizophrenia cognitive problems can lie
phabetical order. The subjects were asked to write down a num- partly in pre-conscious automatic selective attention and partly with the
ber below each word to indicate its original presentation order. speed of processing in later controlled processes (including compound
Subjects were free to work on any list at any time and were allowed STM). The influence of automatic attentional mechanisms may be under-
much time as necessary to complete the task. estimated in normal psychology and explain the loss of the magic 4 in schiz-
Results and discussion. ophrenia.
The relevant data are shown in Figure 1. Each data point rep-
resents average recall performance, collapsed across serial posi- Cowans arguments for a memory store with a capacity limited to 4,
tion, for a given list length. Note that the performance function or thereabouts are persuasive: they are most eloquent in tests of se-
looks very similar to what you would find in a typical short-term rial recall, using articulatory suppression and involving articulated
span experiment. Performance declines with list length in a nearly responses. It is a sort of explicit, short-term memory (STM). To call
linear fashion in fact, a straight line fit of the data accounts for it the focus-of-attention does an excellent service by emphasizing
over 95% of the variance. To calculate long-term memory span, that the store (and by implication attention) are terms and concepts
we simply estimated the list length that produces 50% correct per- applicable not only for information of exogenous origin but also for
formance in this case, the value is 5.15. It is also possible to cal- information with endogenous sources (recall and inter-cortical
culate long-term span when correct recall of the entire list is re- monitoring activities). However, the need to define in and define
quired under these scoring conditions, long-term memory span out certain methodological (e.g., use of 0.5 sec presentation times
drops to 3.75. Both of these estimates are within the range of val- in serial recall, sect. 3.1.1, para. 7) and conceptual considerations
ues typically seen in short-term memory span tasks (Schweickert (e.g., what makes up a chunk? Cf. Shyns et al. 1998) points to some
& Boruff 1986). limits and consequences (see question 4) concerning the imprint of
What do these data mean? The major implication is that a the arrow of information flow on the arrow of time.
chunk-based capacity limit specifically, the magical number 4 The limit of 0.5 sec in tests of serial recall is remarkably conve-
is not a signature characteristic of remembering over the short- nient and reminiscent of Libet et al.s (1964) report that a stimu-
term; as a result, the capacity limitations identified by Cowan say lus should last 500 msec to enter consciousness. Yet, Libet et al.
little, if anything, about whether there is a specialized short-term also remind us that the information is there to be used when it

reaches the cortex after less than 100 msec. Although processing processes (e.g., priming). This suggests to us the prediction that
at this early stage could literally be regarded as implicit, Cowan patients with schizophrenia, renown for their loose associations
suggests that implicit memory stores are also limited to about 4 (see Spitzer 1997) should average a larger capacity than normal:
items. So the sensory buffer or icon remains conceptually intact. Why is there ample evidence that this is not true (e.g., Brebion et
This has been clear, at least since Sperlings (1960) account that al. 2000)? We return to global percepts. These activate the right
cues at short intervals (,1 sec) can elicit correct recall of any of 4 temporo-parietal junction, as emphasised by Frith and Dolan
rows of 4 letters. The buffer may contain up to 16, or even more (1997) in their imaging study, especially during sustained atten-
items. Creation of this buffer also involves selective attention as it tion with few switches of attention. Granholm et al. (1999) found
will occur only if the subject is not concentrating on the experi- not only that processing of local stimuli was impaired in patients
menters tie or distracted by some fluff on the floor. with schizophrenia but (perhaps because of this) under condi-
The separate influences of these 161 and 4-item stores is a mat- tions of divided attention (requiring switches of attention between
ter of daily experience. Faces (250 msec) can be recognised, holis- global and local conditions in the search for a target) were actu-
tically, by an automatic process up to 300 msec later. Only then can ally at an advantage compared to their controls when global pro-
they be decomposed into the elements of eyes, mouth, and so on cessing was required. Switching requires inhibition of the alter-
(ca. 4), by a controlled process (George et al. 2000). Global pro- native, selective attention the inhibition of the irrelevant: its
cessing occurs faster than that for local information, with a peak under-use resulted in the identification of fewer local items. It is
of excitation in the range of 2001 msec, led by temporo-parietal this automatic mechanism of selective attention and its impover-
areas on the right (Sugase et al. 1999; Yamaguchi et al. 2000). The ished use that restricts the patients STM capacity, not just to nor-
efficiency of the early (automatic) selective process, in part, de- mal but below normal levels.
termines the performance of the focus-of-attention. We conclude that the automatic/pre-conscious application of
But we need to look closely at the influences of efficiency and selective attention is not only a source of schizophrenic cognitive
the speed of processing. Increases of the speed of processing im- problems, but has a determining influence on the normal appear-
prove the accuracy of recall and the number of items recalled. This ance of 4 in Cowans pure STM, that he has perhaps underesti-
applies, among controlled conscious processes, to the more su- mated in his worthy review.
perficial encoding of items in pure STM (as in the digit span) as
well as deeper processing in compound STM (as in word lists).
This holds also for patients with schizophrenia (Brebion et al.
2000), for whom there is abundant evidence of impaired STM, If the magical number is 4, how does one
controlled processing and slowed information processing (Straube account for operations within
& Oades 1992). However, it should not be overlooked that con-
trolled selective attention mechanisms are also required, at the working memory?
least to inhibit interfering associations (e.g., Stroop colour-word
Juan Pascual-Leone
interference). It is of interest, here, to note that Brebion and col-
York University, Toronto, Ontario M3J 1P3, Canada. juanpl@yorku.ca
leagues reported that Stroop indicators of selective attention
www.yorku.ca/dept/psych/people/faculty/pasleone.htm
would predict superficial but not the deeper (compound STM)
performance. Yet, Stroop interference performance is not dispro-
Abstract: Cowan fails to obtain a magical number of 7 because his analy-
portionately impaired in schizophrenia (except for those with pre- sis is faulty. This is revealed by an alternative analysis of Cowans own tasks.
dominantly disorganised symptoms), although compound (and The analysis assumes a number 7 for adults, and neoPiagetian mental-
pure) STM measures are reduced in size. This implies that one capacity values for children. Data patterns and proportions of success (re-
problem for patients with schizophrenia lies in making associa- ported in Cowans Figs. 2 and 3) are thus quantitatively explained in detail
tions beyond the temporal and strategic bounds of pure STM and for the first time.
incurs the speed of processing.
But having emphasized the pure STM store and a selective at- Hausman and Wilson (1967) criticized Nelson Goodman (1951)
tention mechanism at the level of controlled processing, we come for his formal-logic model of humans construction of experience.
to the sensory buffer and automatic processing of the through-put He did so because Goodman represented objects (entities) as
to pure STM. While latencies of later event-related potentials collections (sums) of predicates (qualia) that mysteriously com-
(ERPs) such as the P300 (around 400 msec) are usually delayed, bined into objects of phenomenal experience without the need of
the latencies of P50 (marking the thalamo-cortical arrival of in- a combinator or nexus, which integrates this collection into a sin-
formation), N1 (excitatory cortical registration of sensory infor- gle object. This omission, characteristic of empiricist theoreti-
mation) and mismatch negativity (sensory memory for deviance) cians, is harder to imagine today; Computer science and neuro-
are not consistently different among patients with schizophrenia science have recognized the need for a nexus that dynamically
and healthy subjects (e.g., Bender et al. 1999; Gades et al. 1996). combines simple or compound predicates to instantiate objects.
It appears that the speed of automatic (e.g., MMN) as opposed to This nexus problem is currently called the binding problem
controlled processing (e.g., P300) is not delayed, even though the (Kolb & Wishaw 1996; Robin & Holyoak 1995). Cowan acknowl-
content marked by the amplitude of the ERPs recorded is often edges the binding problem, but does not seem to draw its episte-
reduced in both cases. mological implications for task analysis. Indeed, as many do, he es-
Thus it would seem that at short latencies automatic selective timates attentional capacity (working memory) by counting
mechanisms may make a larger contribution to impaired process- distinct schemes that mental-attentional mechanisms must hy-
ing (exaggerated by the state of attention, Oades et al. 1997) than peractivate to cause their coordination. He fails, however, to count
speeds of processing, but the opposite holds for controlled strate- the binding nexus: operative schemes (procedures, combinators)
gic processing. This idea is supported by the finding that the mag- and parameters needed to produce integration or performance
nitude of sensory gating, a selective process, is impaired in schiz- (cf. Robin & Holyoak 1995). Because of this omission, Cowan, like
ophrenia around 100 msec post-stimulus (650 msec: Bender et Halford (1993; Halford et al. 1998; Pascual-Leone 1998) and oth-
al. 1999) and may contribute to the frequently reported phenom- ers, estimates active working memory (attentional capacity) as 3 to
enon of sensory overload. In contrast, local speeds of processing 5 chunks, rejecting Millers insightful number 7.
in different regions (frontal and temporal lobe latencies) may ac- I have analyzed all paradigms that Cowan shows, and results are
count for apparent deficits in the later negative difference consistent with the number 7 when nexus is counted. I now illus-
marker of controlled selective attention (Oades et al. 996). trate these analyses using one paradigm. More refined modeling,
Cowan suggests that the apparent pure-STM capacity can be in- with theoretical probability calculations, can yield more exact
creased by forming inter-chunk relations, perhaps by automatic quantitative predictions confirmed in studies (Burtis 1982; Morra

2000; Pascual-Leone 1970; Pascual-Leone & Baillargeon 1994; These three postulates help to explain data that Cowan reports
Pascual-Leone et al., in preparation). There are also develop- in Figure 2. I discuss only span data. During recall of unattended
mental capacity data of constructivist neoPiagetians, consistent list of digits, in the explicit digit part, adults (on the assumption
with Millers number when extrapolated to adults (Case 1998; that their real attentional capacity is 7) will be very likely to RE-
Case & Okamoto 1996; Johnson et al. 1989; Morra 1994; 2000; CALL 3 digits, as explained above, and RETRIEVE one or two
Pascual-Leone 1970; 1995; Pascual-Leone & Baillargeon 1994; schemes from outside M-space. But against this happening is the
Stewart & Pascual-Leone 1992). fact that overtly recalling three digits that are inside M-space de-
Consider now Cowans experiment reported in section 3.1.3 of mands focussing attention on linguistic schemes of the corre-
the target article. In this experiment, there are two ongoing sub- sponding numerals, and this selective allocation of M capacity
tasks: a visual-rhyme task and an auditory-digit task. One is a com- might lead, with some probability (which increases with the num-
puter game where the picture at the center of the screen tells ber of digits Morra 2000), to interruption/inhibition of digit
which of four surrounding pictures the subject must click; the one schemes that have remained outside M-space. This is what Figure
whose name rhymes with the name of the central picture. New 3 of Cowan confirms: 80% of subjects recall 3 digits, 48.5% recall
items of this subtask are presented while participants hear (but 4 digits, 11.4% recall 5 digits, and only 1 subject (i.e., 2.8%) recalls
must ignore) lists of digits presented through headphones. Occa- 6 digits. Notice that if we assume that 3 is the demand of the ex-
sionally the screen shifts to the digits game, and asks for digit re- plicit visual-rhyme task, 4 digits could still be attended to concur-
call in the appropriate order. Call *pict.cent a subjects figurative rently; and the problem would then be to recall them in order
scheme for the picture at the center, and *pict.surr:x (where x 5 when the unattended-digits task arrives. Ordinal positions of the
1, . . . ,4) the surround picture currently attended to. Call CLICK- first and last digits are easily recalled because of perceptual
MATCH:RHYME the procedure for clicking the matching pic- saliency (this is a known anchor effect); but the relative position of
ture; it has a name-recalling subprocedure (operative parameter) digit 2 and digit 3 would have to be guessed with probability of 1/
that I call NAME, which may be chunked with the first one. Fi- 2 (550% correct), which agrees with Cowan finding 48.5% sub-
nally, the concurrent auditory subtask consists in hearing, without jects recalling 4 digits. If there were 5 digits to be recalled (i.e., 4
intentionally attending, the digit series. Because hearing and vi- from inside M-space plus 1 retrieved from outside M-space), then
sion do not interfere with each other, auditory schemes, released the probability of guessing positions 2, 3, and 4 is 1/6 (6 combina-
by an automatic orienting reaction (OR), produce concurrent im- tions of 3 positions 16 %); close enough to Cowans 11.4%. Finally,
plicit hearing. The eliciting situation is facilitating, because no in- if there are 6 digits available for recall (4 held inside M-space and
terfering auditory scheme is activated by the situation (Pascual- 2 retrieved from outside M-space), the probability of guessing po-
Leone 1987; 1995; in press; Pascual-Leone & Baillargeon 1994). sitions 2 to 5 is 1/24 (4%) again very close to Cowans 2.8%.
Call *digit1, *digit2 . . . *digits, etc., figurative schemes (symbol- The situation is very different in the attended-digits task. In this
ized by a prefixed *) that encode digits being presented; and call task subjects full attention is devoted to digits; and attention may
OR:AUDIT the innately-automatic orienting reaction. With this not be needed to recall digit order because forward serial order
terminology we symbolize; in a mental strategy formula the ex- (natural to language!) was considerably practiced during the span
plicit visual-rhyme part as follows: pretest task (Cowan et al. 1999).
Consequently, when adults are ready to recall, and a RECALL
[CLICKMATCH :RHYME(NAME(*pict.cent, *pict. surr:x)) (1)
operative scheme is placed inside M-space to begin to voice the
& OR:AUDIT(*digit1, *digit2, *digit3, *digit4, *digit5 . . . )]
numerals, 6 digit schemes will also be there. And when the corre-
The first segment has a mental demand of 3 or 4 symbolic schemes sponding 6 numerals are voiced, the digit schemes that are out-
processed simultaneously (it can be 3 when CLICKMATCH: side M-space will be interrupted automatically by virtue of postu-
RHYME and NAME are chunked). The second concurrent seg- late (1) unless adults are sophisticated and postulate (2) applies,
ment is automatic; but on the assumption that adults have a pro- which is unlikely. Consequently, 6 digits should be recalled on av-
cessing capacity of 7 (Pascual-Leone 1970; 1987; 1998; Pascual- erage (see Cowans Fig. 2).
Leone & Baillargeon 1994), we infer that they also allocate The same model explains childrens data, as reported in Figure
attention to 3 digits in this implicit subtask (OR:AUDIT does not 2, when we use theoretical levels of M-capacity predicted for the
need attentional boosting). Thus, when the explicit auditory-digit different age-groups (Pascual-Leone 1970; 1987; Pascual-Leone
part arrives, most adults should remember at least 3 digits. In- & Baillargeon 1994). According to these levels, grade-4 children
deed, Figure 3 of Cowans article shows that 80% of his adult sub- (i.e., 9- and 10-year-olds) have a magical number equal to 4; and
jects do so. But before modeling this explicit auditory-digit part, grade 1 children have a magical number of 3 if they are 7-year-
I will state three principles/postulates of the neoPiagetian con- olds, as they were in Cowans study (1999), who report their mean
structivist theory (Pascual-Leone 1987; in press; Pascual-Leone & age in months and SD). Thus grade-4 children will be able to cope
Baillargeon 1994). Cowans theoretical framework is fully com- easily with the visual-rhyme part of the unattended digits task, and
patible with them (see his whole sect. 2). I list them in 3 points: in addition they may retain 1 digit inside M-space during the vi-
(1) When mental attention (M-capacity) is allocated to schemes sual-rhyme process. Then, during the explicit digit part, children
in the subjects repertoire (making them part of working memory), should retrieve from outside M-space two digits at least.1 The
other activated schemes which are not being attended to (i.e., not task-analysis representation of this step is as follows:
placed in the M-space or focal-attentional region of WM), are gen-
[*digit 1, RETRIEVE(*digit2, *digit3)] (2)
erally automatically inhibited (attentionally interrupted) by the
mental-attentional system; this automatic attentional interruption Thus we expect that grade-4 children will recall, in the unattended
produces the focal beam of attention (Crik 1994; Pascual-Leone digit task, no more than 3 digits on average. Figure 2 shows this to
1987). be the case. As for Cowan s grade-1 children, because their mag-
(2) When the subjects executive processes are sophisticated ical number is 3, they cannot keep any digit inside M-space dur-
enough, task-relevant schemes may be exempted from the atten- ing the visual-rhyme part. Therefore they should recall only digits
tional interruption (automatic inhibition) discussed in point (1). they can retrieve from outside M, that is, usually 2 digits. Figure
(3) Unless they have been so interrupted, some activated cur- 2 confirms this expectation.
rently in an unattended state (i.e., placed outside M-space) can be In the case of the attended digit task, however, grade-4 children
retrieved and brought into focal attention. At least one or two can at first consciously attend to 4 digits, but one must then be
schemes should be so recovered, and, one by one, boosted with placed outside M-space to make room for the operative RECALL.
M-capacity. Notice that recovery is not possible when schemes Then, after the three numerals have been voiced, children are still
have decayed too much or were previously interrupted along with able to retrieve the momentarily dropped (but highly activated)
misleading schemes. digit, plus two other ones from outside M-space to a total of 5

as shown in Figure 2. In the case of Cowans grade-1 children, structure to reflect the properties of the cognitive learning sys-
however, the same argument yields an average number of 4 digits tems likely to be involved. The STM span can be related to lan-
recalled, which is in agreement with data from Figure 2. guage learning. For example, first, it determines the linguistic ma-
The task analysis summarized here illustrates, in a simplified terial that is immediately available to memory for comprehension.
manner, why operative schemes (procedures and their parame- Jarvella (1971) had participants listen to various passages of nor-
ters) should be counted to estimate the demand on working mem- mal discourse, interrupting them at various points to ask how
ory. In doing so, experimental psychologists might find more con- much they could remember of what they had just heard. His main
genial Pascual-Leones (1970) original proposal, which models finding was that there was almost perfect recall for up to seven
developmental emergence of Millers number 7. This proposal of- words before the point of interruption. Second, partly motivated
fers two distinct advantages over the magical-number-4 model of by Newports (1988; 1990) observations in developmental psy-
Cowan and others: (1) It can explain, in a manner congruent with chology, Elman (1993) showed that artificial languages could not
modern neuroscience, the developmental growth of working be learned by a simple recurrent neural network model, unless the
memory from infancy to adulthood, and then its regression in later STM of the model started small and was only gradually in-
years; (2) It can eliminate the many anomalies, in adults working- creased to adult size (to reflect abstraction of regularities of in-
memory performance, that magical-number-4 theoreticians cus- creasing complexity).
tomarily explain away by vaguely appealing to chunking. Indeed, Language learning is likely to be partly mediated via STM (al-
in Cowans paper, I count 14 occasions when the authors scrupu- ternatively, language is likely to have co-evolved with STM). Thus,
lous review acknowledges research results that suggest a magic we would expect the language statistical structure to reflect prop-
estimate of 6 or 7 but chunking can be used to explain this inter- erties of STM, and, in particular, STM size. This is not to say that
pretation away. I discuss this problem further in a BBS commen- we cannot process linguistic contingencies that exceed STM size,
tary (Pascual-Leone 1998) that complements the present one. but if the cognitive system is optimized to process automatically
statistical associations only within a certain range (namely STM
ACKNOWLEDGMENTS span), we would likewise expect language structure to be consis-
Research informing methods and ideas summarized in this commentary tent with this limitation.
were supported by Operating Research Grants from the Social Sciences Cowan presents a compelling line of evidence to argue for an
and Humanities Research Council of Canada. STM span range between three and five chunks. Consistent with
I am grateful to Dr. Janice Johnson, Dr. Sergio Morra, and Antonio this observation and relevant to language learning, Cleeremans
Pascual-Leone for insightful comments that found their way into the final and McClelland (1991) showed that in learning regularities in se-
manuscript. quentially presented material, participants could not identify con-
NOTE
tingencies involving stimuli separated by more than about three
1. We expect that digit-schemes from outside M-space will be recalled other stimuli.
with higher probability whenever shorter lists are presented, because in- In this work we look at the statistical correlational properties in
terference among digit-schemes increases with their number; and with text from eight languages. On the basis of the above finding that
this number so does the possibility of error in encoding and retrieval op- language structure involves contingencies between word tokens
erations. For this reason, and because the length of lists changes with age that are separated by about four other word tokens would provide
in Cowans study, we expect that younger children recall more digits from additional support for Cowans proposal in the important area of
outside M-space than adults. language learning and processing.
Mutual information and linguistic structure. All the analyses
presented are based on the notion of mutual information (MI), a
measure of relatedness between different probability distribu-
Linguistic structure and short term memory tions. Range is defined as the number of words between two
given word tokens, x and y. For instance, a range of 1 will indicate
Emmanuel M. Pothos and Patrick Juola that words x and y are separated by only 1 other word; word to-
aSchool of Psychology, University of Bangor, Bangor LL57 2DG, United kens x and y are separated by range 0 if and only if x and y are ad-
Kingdom; bDepartment of Mathematics and Computer Science, Duquesne jacent. We ask whether our expectation of obtaining word y at a
University, Pittsburgh, PA 15282. e.pothos@bangor.ac.uk particular location is affected by the knowledge that we have word
juola@mathcs.duq.edu www.bangor.ac.uk/~pss41b x in an earlier location. A measure of this is the mutual informa-
www.mathcs.duq.edu/~juola tion (MI) between P(x) and P(y), the probabilities of obtaining
word x and word y respectively. Mutual information indicates how
Abstract: We provide additional support for Cowans claim that short term much the uncertainty involved in predicting y is reduced by knowl-
memory (STM) involves a range of 3 5 tokens, on the basis of language edge that we have x, and is given by Sum [all x, y] (P(x, y)/(P(x)/
correlational analyses. If language is at least partly learned, linguistic de-
pendency structure should reflect properties of the cognitive components
P(y)). For different ranges, P(x, y) is the probability of having
mediating learning; one such component is STM. In this view, the range words x and y separated by a number of words equal to the range.
over which statistical regularity extends in ordinary text would be sugges-
tive of STM span. Our analyses of eight languages are consistent with STM
span being about four chunks. Table 1 (Pothos & Juola). SE is the standard error of the
mean sample size, for each language
Introduction. We take the view that language learning involves
a major component of automatically identifying regularities in se- Language Mean Words SE Samples
quentially presented material. While there has been an influential
research tradition arguing that language learning must be guided Bulgarian 1,468 256 4
by a substantial innate component (Chomsky 1975; Pinker 1994; Czechoslovakian 27,591 860 29
see Gold 1967 or Pinker 1979 for a mathematical analysis of this Dutch 181,407 33,483 35
problem and Wharton 1974 for important qualifications on Golds
English 97,272 11,805 12
results), considerable evidence has accumulated recently that lan-
guage can be largely inferred on the basis of experience (Elman Estonian 19,944 15,043 2
1996; reviews of statistical models of language and, in particular, French 166,620 202 26
connectionist ones are given respectively in Charniak 1993, and Gaelic 200,239 1
Chater & Christiansen 1999). German 129,270 96,532 8
If language is (at least partly) learned, then we expect language

cal accounts often mention the role that could be played by neural
oscillations and synchrony. For example, Lisman and Idiarts (1995)
model might account for a four-item limit of short-term memory,
given an appropriate frequency ratio of nested oscillations.
As an alternative, we present a recently developed cortical net-
work model accounting for the limited capacity of visual working
memory (Raffone & Wolters, in press). A critical capacity of about
four independent patterns showed up in our simulations, consis-
tent with the results of Luck and Vogel (1997) and with Cowans
conclusions. This capacity was proven to be independent of the
number of features making up objects (Luck & Vogel 1997). The
model mechanism may be applicable to any type of information in
working memory, and not only to visual information. In the fol-
lowing we present some details of this model, and relate it to some
of the functional aspects of working memory discussed by Cowan.
Retention in the cortical circuits of working memory. In the
model, we assume that the cortical circuits of visual working mem-
ory are implemented in prefrontal (PF) cortex and visuo-tempo-
ral areas (e.g., the inferotemporal cortex, IT), as well as in their
mutual connections. Oscillations were induced by external input
Figure 1 (Pothos & Juola). Mutual information profile when MI and maintained after input offset through delayed feedback from
values were standardized for the different languages. PF to IT (see Fig. 1A) In Lisman and Idiarts model persistent fir-
ing is due to a slowly rising after-depolarization (ADP) state in
combination with a sub-threshold oscillatory modulation in the
By MI profile, we mean the way MI varies with increasing range. theta range. In our view, reverberatory oscillations are more likely
This we take to be an indicator of statistical structure in language. to occur as a result of cooperative interactions among many neu-
Analyses. We investigated samples from eight different lan- rons, involving both prefrontal assemblies and assemblies located
guages, all from the CD-ROM database of the European Corpus in higher-level perceptual areas.
Initiative Multilingual Corpus 1 (ECI/MCl), distributed by the Between-item segregation in working memory. Assuming that
Association for Computational Linguistics. In all analyses, the lin- time-resolved neural coding plays a key role in (visual) working
guistic tokens examined are words (however, the same type of memory processing, the limited capacity of working memory may
investigation can be conducted on, e.g., phonemes; see Pothos be explained in terms of spurious synchronizations of neuronal
1998). Table 1 shows the number of samples and average number spike trains coding different unrelated features or items (Luck &
of words in each language. Vogel 1997). It is unclear, however, how such spurious synchro-
The MI profiles for each language were averaged and standard- nizations of to-be-retained features could give rise to critical ca-
ized; standardization enables us to compare MI values for each lan- pacity limits, even if averaged across trials and subjects. We be-
guage regardless of sample size differences. Figure 1 shows the re- lieve a more satisfactory explanation of the limited capacity of
sults of this calculation. One can see that the mutual information visual working memory is provided by a between-item segregation
dependence elbows at about four items for all the languages. We hypothesis. According to this hypothesis, neural assemblies in
take this observation to be compatible with the idea that linguistic high-level visual areas, coding unrelated features or objects, exert
dependency structure is contained primarily within a range of not mutual inhibitory or desynchronizing actions. The network auto-
more than about five word tokens, which would be consistent with matically scales the phase-lag between the different reverbera-
the capacity of STM span being of the order of 35 chunks. This tions in order to maintain an optimal phase segregation (see Fig.
provides an additional source of evidence for Cowans proposal. 1B). Between-item segregation depends on the inhibitory synap-
With additional work we aim to extend the present MI computa- tic parameters. The essence of the model is that its limited capac-
tions to take into account more directly sample size variation, and ity depends on the functional balance of a sufficiently high oscil-
also introduce a mathematical model for the MI profiles. lation frequency (firing rate) and sufficient phase-segregation
between disjoint assemblies. The same desynchronizing mecha-
nism optimizing phase-segregation between assemblies coding for
separate items, poses a limit to the number of oscillatory rever-
berations. Thus, our model suggests plausible neurocomputa-
A neurophysioiogical account of working tional reasons of the short-term memory (STM) capacity limit. As
memory limits: Between-item segregation the model operates in a stochastic manner, it exhibits a psycho-
and within-chunk integration logically plausible variability across trials of the number of objects
retained, given the same number of to-be-retained objects as in-
Antonino Raffone,a Gezinus Wolters,b and Jacob M. Murrec put.
a
Department of Psychology, University of Rome La Sapienza, I 00185 Within-chunk integration in working memory. As is clear from
Rome, Italy; bDepartment of Psychology, Leiden University, 2300 RB Leiden, Cowans discussions, the notion of a capacity limitation is inextri-
The Netherlands; cDepartment of Psychology, University of Amsterdam, 1010 cably interwoven with the concept of chunking. In our view,
WB Amsterdam, The Netherlands. raffone@uniromal.it
chunking is related to neural binding processes, which may origi-
wolters@fsw.leidenuniv.nl jaap@murre.org
nate from bottom-up attentive operations or may be guided by
long-term memory. Following the suggestion of Luck and Vogel
Abstract: We suggest a neurophysiological account of the short-term
memory capacity limit based on a model of visual working memory (Raf-
(1997), we modeled intra-chunk integration or binding in short-
fone & Wolters, in press). Simulations have revealed a critical capacity term memory retention in terms of synchrony. Chunking may be
limit of about four independent patterns. The model mechanisms may be based on pre-existing neural assemblies (chunking fields), with
applicable to working memory in general and they allow a reinterpretation synchronizing connections between assemblies coding bound fea-
of some of the issues discussed by Cowan. tures, or on correlated input from earlier processing stages.
In the model, within chunk integration coexists with between-
Why the capacity limit? As Cowan points out, the reasons for a chunk segregation (see Fig. 1C). Such a property cannot be ac-
short-term memory capacity limit are not clear. Neurophysiologi- counted for by delay activity defined in terms of firing rate attrac-


tors (Amit 1995). However, the network capacity for single- and anisms. He suggests (sect. 3.1.1) that a similar limit may apply to
multiple-feature chunks is equivalent only when there is a high change blindness, the finding that large changes become difficult
chunk stability. This functional scheme can potentially account to see when information about the location of the change is
for the different degrees of chunking effectiveness in short-term swamped by concurrent transients elsewhere in the visual field
memory retention discussed by Cowan. (Rensink et al. 1997; Simons 1996).
Central capacity versus separate capacities. In several places In a typical change-blindness experiment, an original and a
Cowan discusses the central versus separate capacity contro- modified image of a real-world scene are presented in succession,
versy. Our proposal can account for various degrees of domain with a brief blank field between them; alternation continues until
specificity in processing limitations, by the assumption that com- observer detects the change. Even though the changes are large
petition and desynchronizing actions between neural assemblies and the observer knows they will occur, several seconds are often
coding for unrelated features are stronger within than between required before a change is seen. This has been explained by the
specific representational domains. hypothesis that focused attention is needed to see change (Ren-
Storage versus processing in working memory. In section sink et al. 1997). Given that focused attention can be largely iden-
4.3.6, Cowan discusses the storage versus processing capacity ac- tified with visual STM (vSTM), and that vSTM has a limited ca-
count, concluding that there is no reason for a separation between pacity, only a few items can be attended at any time. Thus, the
processing and storage in working memory. We fully agree with detection of change requires a time consuming attentional scan of
this view. In the cerebral cortex, a bi-directional interaction be- the image.
tween short-term reverberations and long-term structured neural But how many items can be attended at any one time? (Or,
assemblies is likely to occur. In such a scheme, time-resolved syn- equivalently, how many can be held in vSTM?) This can be deter-
chronization (related to firing rate enhancement) and Hebbian mined from change-detection experiments based on arrays of sim-
learning may functionally interact. ple items (Rensink 2000a). The critical parameter here is on-time
The focus of attention. Cowans proposal of a central capacity (the length of time each array is visible during a cycle). The time
limit of the focus of attention leads to some neurophysiological needed to detect a changing target item among nonchanging dis-
problems. That is, how does a mechanism both collect and read tractors depends linearly on the number of items in the display.
out information from other sub-systems? Is it centralized or distrib- For orientation change, the slope of this function (i.e., search
uted in structure? How does it access and operate on the distrib- speed) is much the same for all on-times up to 600 msec, indicat-
uted representations in long-term memory? Cowan repeatedly ing that the rate-limiting step is one of processing rather than
refers to the widely shared idea of activation of memory repre- memory. But for on-times of more than 600 msec, speed becomes
sentations. However, such an idea can lead to coding problems proportional to alternation rate, indicating that only a limited
when multiple distributed neural representations are active, lead- amount of information can be held in vSTM at each alternation
ing to the so-called superposition catastrophe (Engel et al. 1992; more display time does not allow more items to be entered into
Von der Malsburg 1901). We suggest that a neural mechanism memory. When the interstimulus interval (ISI) between displays is
based on both activation (firing rate) and coherence (neural syn- 120 msec, this limit is 5 6 items (Rensink 2000a). Further exper-
chrony) is more plausible for the central read-out bottleneck. iments have shown this to be a compound limit: when a short-term
We endorse Varelas notion of resonant assemblies in the brain presumably iconic component is eliminated by increasing ISI
(Varela 1995). In terms of our model, this would imply that only to 360 msec, the estimate falls to 3 4 items (Rensink et al. 2000).
four magical neural assemblies can dominate the brain at any As Cowan points out, it is important to establish the absence of
given time, saving other domain-specific active neural represen- rehearsal or recoding processes that might cause estimates to be
tations through selective synchronization. artificially high. For change detection, this is straightforward.
First, the situation is one of information overload: not all the visi-
ble items can be placed into memory. Second, little recoding or re-
hearsal can occur (at least for cycle times of a second or less), since
most of the available time is spent either loading items into mem-
Four-sight in hindsight: The existence ory or comparing them with the current input. Third, capacity is
of magical numbers in vision determined by a genuine discontinuity in performance, namely, a
proportionality constant that appears when on-times are 600 msec
Ronald A. Rensink or greater (Rensink 2000a). Finally, the estimate is largely unaf-
Cambridge Basic Research, Nissan Technical Center North America, Inc., fected by temporal decay: if ISI is greater than 360 msec, there is
Cambridge, MA 02171-1494. rensink@cbr.com
little further decrease, even for intervals as high as 8 sec (Schnei-
www.cbr.com/~rensink/
der et al. 1999). Thus, the magical number 4 does seem to exist.
But the story does not stop here. If targets and distractors are
Abstract: The capacity of visual attention/STM can be determined by
change-detection experiments. Detecting the presence of change leads to
switched so that the subject must detect a nonchanging target
an estimate of 4 items, while detecting the absence of change leads to an among changing distractors, a different limit is reached: 1.4 items
estimate of 1 item. Thus, there are two magical numbers in vision: 4 and (Rensink 1999; 2000b). This suggests that attended items are not
1. The underlying limits, however, are not necessarily those of central independent but are instead pooled into a single collection point,
STM. or nexus (Rensink 2000b). Such a magical number 1 may corre-
spond to the limit alluded to by Cowan in his proposal that the [4
In his target article, Cowan provides a wide-ranging review of data separate] parts are associated with a common higher-level node
supporting the existence of a magical number 4 a common (sect. 2.6).
limit on the capacities of various perceptual and cognitive mech- Therefore, there is considerable support for the claim of at least
Figure 1 (Raffone et al.). (A) Individual assembly dynamic behavior with feature input in IT and active feedback from PF. The panel
shows the evolution and continuation of the average activity of one IT assembly (100 interconnected model neurons coding a single fea-
ture). Stimulus onset and offset times are marked by the vertical lines. (B) Phase segregation of IT assemblies coding for disjoint fea-
tures. Four out of five reverberations remain active. Due to mutual inhibitory activity, the assemblies become spaced in the oscillatory
phase, thus allowing a markedly discriminative oscillatory reverberation and retention of the coded features. Assemblies are shown in an
order allowing easy inspection of phase segregation. (C) The combination of within-chunk integration and between-item segregation.
Objects (chunks) consist of four interconnected assemblies. Four out of five objects are retained in terms of internally synchronized and
mutually desynchronized oscillatory chunks, whereas all features coding a fifth object are suppressed.

two magical numbers in vision. However, there is less support for The scope of base groupings has to be limited. It is unlikely that
the claim that these are due entirely to limitations on a central there are cells in higher visual areas that are tuned to arbitrary fea-
working memory. To begin with, the long-term memory units (or ture constellations (von der Malsburg 1995). This implies that an
chunks) accessed by STM need not be the same as the vSTM units additional type of grouping, called incremental grouping, is re-
(or parts) obtained from the visual input. A particular configura- quired if base grouping fails to do the job. This may be necessary
tion might be a unit for purposes of memory retrieval, but not for if relationships need to be established between feature domains
visual operations such as tracking or attentional suppression. Dif- that are not available as base groupings. In this case separate neu-
ferent kinds of processes are likely to be involved, and thus, dif- ronal populations encode the different feature domains, and the
ferent kinds of units. binding problem lurks. To avoid binding problems, assemblies of
As an illustration of this, consider the detection of change in neurons that respond to the various features of the same percep-
contrast sign. Whereas capacity for orientation is 3 4 items, ca- tual group have to be demarcated from other neurons that re-
pacity for contrast sign is at least 10 items (Rensink 2000a; 2000c). spond to different groups. This can be achieved by the use of as-
This is likely to be a compound limit, resulting from the grouping sembly labels that are shared among neurons belonging to the
of items of similar contrast sign. But note that such groups are same assembly. Previous theories have proposed two possible as-
purely short-term visual structures there is little likelihood that sembly labels: synchrony and firing rate modulations (reviewed by
any particular arrangement had been seen before and became a Roelfsema & Singer 1998). According to the first possibility, neu-
chunk in long-term memory. rons that belong to the same assembly fire in synchrony. The sec-
More generally, perception and cognition rely on systems which ond possibility is that neurons of the same assembly are labeled
interact with each other to a high degree, making it difficult to de- with an enhanced firing rate. The distribution of either of these
termine the locus of performance limits. Indeed, there may not assembly labels across neurons takes time, and incremental group-
even be a single locus: performance on a visual task may involve ings therefore do not form automatically (Roelfsema et al. 2000).
both visual structures (parts) and memoric structures (chunks); a Let us now address the limits that are imposed by these assembly
magical number might represent the number of degrees of free- labels on the number of objects that can be stored in STM.
dom on a structure linking the two levels. Given that there are no When synchrony is used as an assembly label, multiple assem-
compelling a priori grounds which can be appealed to, this mat- blies can coexist. Neurons within each of these assemblies may fire
ter will have to be settled by experiment. (The issue of individual synchronously, but they need not be synchronized to neurons of
differences would seem to be a particularly good candidate in this the other assemblies. Thus, multiple incremental groupings can
regard.) Until such experiments are carried out, it may be best to form during stimulus presentation and they can, in principle, also
keep our options open as to what causes the magic in our visual be stored in STM. The target article embraces the synchrony la-
world. bel and suggests that the number 4 emerges as the ratio between
brain rhythms, such as the 40 Hz gamma rhythm, and the 10 Hz
alpha rhythm, as was first suggested by Lisman and Idiart (1995).
The underlying idea is that four assemblies that are synchronized
Which brain mechanism cannot at the higher frequency can just stay out of phase in a single cycle
of the low frequency oscillation. However, this line of reasoning
count beyond four? neglects another result of studies on cortical synchronization.
Pieter R. Roelfsemaa and Victor A. F. Lammea,b Neurons that respond to different objects do not have a tendency
aGraduate School Neurosciences Amsterdam, Department of Visual System
to stay out of phase. Instead, neurons that respond to different ob-
Analysis, Academic Medical Center (UvA), 1100AA Amsterdam, The
jects usually have weaker synchronization, or fire independently
Netherlands; bThe Netherlands Ophthalmic Research Institute, Amsterdam, (Gray et al. 1989; Livingstone et al. 1995). This invalidates argu-
The Netherlands. p.roelfsema@ioi.knaw.nl v.lamme@amc.uva.nl ments that are based on the ratios between brain rhythms (e.g.,
www.ioi.knaw.nl/vsa.iwo Lisman & Idiart 1995). Because different assemblies do not stay
out of phase, spurious synchronization between cells that belong
Abstract: Cowan makes an intriguing case for a fundamental limit in the to different assemblies will occur. This may also account for the
number of chunks that can be stored in short term memory (STM). interference between multiple items that are stored in STM.
Chunks are collections of concepts that have strong associations to one an- However, in this case the number 4 does not immediately fall out
other and much weaker associations to other chunks. A translation of this of the theory.
definition for the visual domain would be that a visual chunk is a collec- The other label that can be used to demarcate neurons that re-
tion of features that belong to the same perceptual group (see also Ma- spond to the various features of a single object is an enhancement
honey & Ullman 1988). Here, we will first address the neuronal mecha-
nisms that may demarcate visual chunks. Then we critically evaluate to
of their firing rate. Such response enhancements have often been
what extent these mechanisms might be responsible for the limit on the documented as the correlates of visual attention that is directed to
number of chunks that can be held in STM. We conclude that the clarity the respective object (reviewed by Desimone & Duncan 1995;
with which the psychophysical data point to the number four is not Maunsell 1995). Grouping of features by attention has first been
matched by a similarly clear limit imposed by physiological mechanisms. suggested in Treismans feature integration theory (e.g., Treisman
& Gelade 1980). The enhanced firing rate label can, however, only
It is important to distinguish between two types of grouping: base be used by one assembly at a time. If two assemblies responding
grouping and incremental grouping (Roelfsema et al. 2000). Base to different objects are simultaneously labeled, the binding prob-
groupings are formed rapidly and automatically after the appear- lem reappears, because neurons in both assemblies have an en-
ance of a novel image, because they are based on the tuning of in- hanced firing rate. This implies that rate modulations only allow a
dividual neurons. Many single neurons in early (e.g., Leventhal et single incremental grouping at a time (Roelfsema & Singer 1998).
al. 1995) as well as in higher visual areas (e.g., Kobatake & Tanaka What happens to base- and incremental groupings when the
1994) are tuned to multiple stimulus attributes, such as color, ori- stimulus is removed and bottom-up activation declines? The inti-
entation, and motion direction. The activation of a neuron tuned mate relationship between STM and attention that is discussed in
to, for example, red and vertical provides a base grouping between the target article, is also present at the physiological level. For
these features. Most neurons in all of these visual areas are acti- many neurons, the enhancement of neural responses to attended
vated within 100 msec after stimulus presentation (e.g., Oram & features persists during intervals in which the visual stimulus is re-
Perret 1992). Thus, base groupings are rapidly available, and do moved (Chelazzi et al. 1993; Fuster 1997; Rainer et al. 1998).
not depend on elaborate processing. There is no clear limit on the Thus, the very same neurons that carry the attentional label are
number of base groupings that can be computed in parallel dur- also involved in STM. The largely unexplored mechanisms that
ing stimulus presentation. are responsible for persistent firing during memory episodes may

impose a limit on the number of neurons that can maintain their

activity. It is unclear, at present, whether such a limit would affect
the number of base groupings or rather the number of incremen-
tal groupings that can be stored. In any case, the number 4 also
does not seem to fall out of the theory here.
In summary, synchrony allows the coexistence of multiple in-
cremental groupings, whereas rate modulations only accommo-
date a single incremental grouping at a time. Four is not a clear-
cut prediction of either mechanism. The data of Luck and Vogel
(1997) show that feature conjunctions of 4 objects are held in
STM. It is, however, left open by this study whether these con-
junctions go beyond base grouping, since individual neurons in
early visual areas can also be tuned in multiple feature domains.
Moreover, the four visual objects that are held in STM cannot have
arbitrary complexity. Thus an exciting arena for future psy-
chophysical and physiological experimentation is defined, to in-
vestigate how many incremental groupings can be held in visual
STM.
Functional neuroimaging of short-term

memory: The neural mechanisms
of mental storage Figure 1 (Rypma & Gabrieli). Two axial (horizontal) slices show-
ing averaged neural activity across all subjects in the 3-and 6-
Bart Rypmaa and John D.E. Gabrielib letter memory load conditions. A 5 anterior, P 5 posterior, L 5
a
University of California, Department of Psychology, Berkeley, CA 94720- left, R 5 right. Gray arrows indicate dorsolateral PFC, white ar-
1650; bDepartment of Psychology, Stanford University, Stanford, CA 19001. rows indicate ventrolateral PFC. FMRI results indicated ventro-
rypma@socrates.berkeley.edu gabrieli@psych.stanford.edu lateral PFC activity in the 3-letter condition and additional dorso-
lateral PFC activity in the 6-letter condition.
Abstract: Cowan argues that the true short-term memory (STM) capac-
ity limit is about 4 items. Functional neuroimaging data converge with this
conclusion, indicating distinct neural activity patterns depending on
cortex (PFC), the only observable difference was increased activ-
whether or not memory task-demands exceed this limit. STM for verbal
information within that capacity invokes focal prefrontal cortical activation ity for the 3-letter load in inferior (ventrolateral) parts of the left
that increases with memory load. STM for verbal information exceeding hemisphere (around Brocas area) (Fig. 1). The second fMRI
that capacity invokes widespread prefrontal activation in regions associ- analysis examined the difference between 6-letter and 1-letter
ated with executive and attentional processes that may mediate chunking loads, a difference that not only increased STM demand, but also
processes to accommodate STM capacity limits. exceeded capacity limits, thus invoking executive processes. In-
deed, there was a dramatic activation increase, not only around
Cowan provides a thoughtful integration of broad literature sup- Brocas area, but in many additional areas, including superior (dor-
porting the view that: (1) 4 6 1 represents the true capacity limit solateral) regions of the PFC, in both hemispheres, and in cingu-
of STM, and (2) supracapacity maintenance (i.e., more than 45 late gyrus. Altogether, there was a 75-fold increase in the spatial
items) of information occurs as a result of additional executive 2 extent of PFC activation as subjects moved from subcapacity to
processes that expand capacity by strategic data-compression (i.e., supracapacity SIM performance. Thus, during supracapacity mem-
intelligent grouping or chunking e.g., Baddeley 1986; Waugh ory performance, we observed a greatly disproportionate increase
& Norman 1965) of to-be-maintained information. This funda- in PFC activation, and activation of multiple PFC regions.
mental idea about the structure of the human mind is, however, These results can be readily interpreted to support the view
controversial because, as Cowan reviews, alternative interpreta- that distinct STM processes, mediated by separable neural sys-
tions of relevant experimental results are possible. tems, carry out information retention, depending on the extent of
Cognitive neuroscience can offer independent and convergent memory-demand (i.e., whether or not the memory demand ex-
constraints upon such a theory of STM functional architecture. ceeds a 4 6 1 capacity limitation). In PFC, variation within that
One functional neuroimaging study that we performed (Rypma et capacity (1 vs. 3 letters) resulted in focal left ventrolateral activa-
al. 1999) offers remarkable support for Cowans argument. tion that increased with load. This activation may reflect a STM
That study examined the neural correlates of STM with func- system that mediates subcapacity maintenance (i.e., less than 4
tional magnetic resonance imaging (fMRI), a neuroimaging tech- items) of verbal information, possibly through verbal rehearsal
nique that allows detection of hemodynamic changes associated (e.g., Vallar & Baddeley 1984). STM loads exceeding that capac-
with regional neural activity that accompanies cognitive perfor- ity (6 letters) invoked widespread bilateral dorsolateral PFC acti-
mance. Subjects performed a verbal STM task, based on Stern- vation in areas not activated by subcapacity STM loads. These ac-
bergs (1969) item-recognition task, in which STM load was var- tivations may reflect the additional executive involvement in
ied from 1 to 3 to 6 letters. Thus, the 1-letter and 3-letter loads fell memory maintenance tasks with supracapacity memory loads dis-
within the putative true capacity of STM, whereas the 6-letter load cussed by Cowan (e.g., Baddeley & Hitch 1974) Indeed, multiple
exceeded that capacity. Subjects saw memory-sets of 1, 3, or 6 let- studies suggest that dorsolateral PFC is activated whenever exec-
ters per trial, kept those letters in mind for 5 seconds, and then utive processes are required (Cohen et al. 1994; DEsposito et al.
pressed a key, if a probe-letter was part of the memory-set. As is 1995; Petrides 1996; Prabhakaran et al. 2000; Smith & Jonides
typically found, subjects were highly accurate and reaction times 1999). Further, concomitant activation in dorsolateral PFC and
increased linearly with increasing memory load. cingulate gyrus have been observed in attention-demanding tasks
One fMRI analysis examined the difference between 3-letter (e.g., Corbetta et al. 1991), supporting Cowans contention that at-
and 1-letter loads, a difference that may be conceptualized as an tention is one component of the additional mechanisms that sup-
increase in STM demand, but within capacity limits. In prefrontal port chunking.

The behavioral results reviewed by Cowan and our fMRI results limit of (STM) is about 4 chunks, rather than the classical 7 6 2.
together indicate that verbal STM loads below a 4 6 1 capacity Concerning visual working memory, this capacity estimate has in-
limit invoke a specific STM system mediated in part by left infe- deed been put forward by several authors for quite some time.
rior PFC. Supracapacity verbal STM loads invoke multiple addi- Based on experimental evidence from studies on transsaccadic
tional executive and attentional systems mediated by bilateral dor- memory (e.g., Irwin 1992) and on visual STM (e.g., Shibuya &
solateral prefrontal and cingulate cortices. STM circuitry extends Bundesen 1988), one of us reached a similar conclusion in a the-
beyond these areas to at least parietal and cerebellar regions, but oretical analysis of visual working memory (Schneider 1999).
these prefrontal findings offer independent, convergent evidence However, knowing how many chunks can be retained is by no
for Cowans persuasive argument. way sufficient if we want to know how much information can be
stored in STM to answer this question, it is absolutely necessary
to know more precisely how a visual chunk can be characterized.
Indeed, we think that a central and important issue of future re-
Characterizing chunks in visual short-term search will be to analyse the limits of STM in terms of the mech-
anisms by which more basic elements are formed into chunks, and
memory: Not more than one feature also to study how attention and specific tasks determine this
per dimension? chunking.
Imagine a scene consisting of four objects, each being a
Werner X. Schneider, Heiner Deubel, square made up of a 3*3 raster of differently coloured parts. An
and Maria-Barbara Wesenick intuitively obvious prediction would be that these four objects can-
Institute of Psychology, Ludwig-Maximilians-University Munich, D-80802 not be retained as well as four homogeneously coloured squares.
Munich, Germany. {wxs; deubel; wesenick}@psy.uni-muenchen.de Amazingly, the data from one of Luck and Vogels (1997) experi-
www.paed.uni-muenchen.de/mip/psych/deubel/wwwdocs/index.htm
ments seemed to suggest otherwise. In this experiment, subjects
had to retain such stimuli as shown in Figure 1 that were defined
Abstract: Cowan defines a chunk as a collection of concepts that have
by various colour-colour-conjunctions. The results seemed to
strong associations to one another and much weaker associations to other
chunks currently in use. This definition does not impose any constraints show that objects that are made of a conjunction of two colours
on the nature and number of elements that can be bound into a chunk. We could be retained equally well as objects with just one colour value
present an experiment to demonstrate that such limitations exist for visual per object. This implies that visual chunks may contain at least two
short-term memory, and that their analysis may lead to important insights feature values for the colour dimension.
into properties of visual memory. For storage of visual information such an analysis is possibly
somewhat easier than for verbal material, which is the focus of
To determine the capacity limit of short-term memory (STM) can Cowans article. In the visual format it is intuitively plausible to as-
be a tricky business with a number of potential pitfalls. Cowan sume that chunks are direct reflections of consciously perceived
provides a careful and much needed analysis of these pitfalls that visual objects. A frequent assumption in vision is that an object
can confound estimates of the real memory capacity with other can be described in terms of its basic visual dimensions, such as its
limitations in the system. After reviewing a large number of stud- colour, shape, or motion. Within each dimension, the stimulus can
ies that try to avoid these problems, Cowan concludes that the be specified by features or feature values (e.g., Treisman &
Figure 1 (Schneider et al.). Left: Stimuli used in the experiment. Subjects had to memorize configurations of 2, 4, or 6 coloured squares
for 900 msec. After the retention period, either the same display or a display with one of the squares differing in colour reappeared; sub-
jects were asked whether they had detected a change. Squares could be large (0.75*0.75 deg) or small (0.35*0.35 deg), or conjunctions
of both. Right: Percent correct report as a function of the number of items presented.

Gormican 1988) For instance, the dimension colour of an object The magical number 4 in vision
can be characterized by one of the features red, blue, purple,
and so on. After breaking down a visual object into these more Brian J. Scholla and Yaoda Xub
basic elements, the question arises as to whether the above defi- a
Department of Psychology, Vision Sciences Laboratory, Harvard University,
nitions of dimensions and features in vision constitute an ap- Cambridge, MA 02138; bDepartment of Brain and Cognitive Sciences, Mas-
propriate basis for characterizing chunks in visual memory. In sachusetts Institute of Technology, Cambridge, MA 02139.
other words, one may ask (1) whether chunking in visual working scholl@wjh.harvard.edu yaoda@psyche.mit.edu
www.wjh.harvard.edu/~scholl
indeed functions by binding features into integrated visual ob-
jects, and (2) analyse the rules and limitations of this binding.
With regard to the first question there is recent evidence from Abstract: Some of the evidence for a magical number 4 has come from
Luck and Vogel (1997, also referred to in sect. 3.1.1), suggesting the study of visual cognition, and Cowan reinterprets such evidence in
terms of a single general limit on memory and attention. We evaluate this
that the capacity limit of visual STM indeed refers to feature bun- evidence, including some studies not mentioned by Cowan, and argue that
dles in the form of objects. In their experiments, subjects were re- limitations in visual processing are distinct from those involved in other
quired to retain simple geometrical visual objects made up of fea- memory phenomena.
ture conjunctions such as of a certain colour, orientation, and
length. The data showed that objects defined by conjunctions of Cowans discussion of the magical number 4 synthesizes evidence
two or more dimensions (e.g., a line of a certain orientation, from domains which are rarely discussed together. In particular,
colour, and length) can be retained as well as objects defined by Cowan draws on work from the study of visual cognition such as
only a single dimension (e.g., orientation only). For any of these studies of subitizing (sect. 3.3.2) and multiple object tracking
combinations, the estimated memory capacity was about four ob- (sect. 3.3.3) and attempts to reinterpret such evidence in terms of
jects. a general memory limitation, which he suggests is a reflection of
The second question is strongly related to the problem of how the underlying capacity of the attentional focus (a thesis which
many of these basic elements can be bound into a single chunk, is discussed in Cowan 1995, but which he does not argue for in his
and whether there exist limitations as to the possible combination target article). Here we note additional evidence for a limit of ap-
of features. proximately 4 objects in certain types of visual processing, and dis-
In our own experiments, we attempted to replicate this sur- cuss why these limits are probably distinct from those involved in
prising finding (Deubel et al., in preparation). We used the same other (e.g., verbal) tasks.
stimuli as Luck and Vogel (1997), and identical experimental pa- Additional evidence from visual cognition. Additional evi-
rameters such as presentation and retention times. Our experi- dence for a magical number 4 in visual processing comes from
mental results (Fig. 1) clearly show that retention of objects de- studies of infants, normal adults, and neuropsychological syn-
fined by a conjunction of two colours leads to a strong drop in dromes. Recent looking-time studies with infants have suggested
performance, as compared to the condition in which the objects that they are able to keep track of arrays of objects through addi-
consisted of one colour only. So, external objects with two colours tions and subtractions, but only if there are less than 4 objects in
seem to require two chunks for the internal coding. This finding these arrays (e.g., Uller et al. 1999; Wynn 1992), and this evidence
is in obvious contrast to the result of Luck and Vogel (1997). The has been interpreted in terms of developing mechanisms of visual
reason why we could not replicate their data is unclear to us, how- attention (e.g., Carey & Xu, in press; Scholl & Leslie 1999). In nor-
ever, in an independent study, Wheeler and Treisman (submitted) mal adults, there appears to be a limit of 4 on the number of ob-
recently reported a finding similar to ours. jects which can receive prioritized processing due to attentional
These data are clear evidence that there exist prominent limi- capture (Yantis & Johnson 1990), and the number of items which
tations to chunking in visual memory. As a possible, preliminary can be simultaneously examined in a visual search for a change
rule of thumb suggested by the result, one may assume that a vi- (Rensink 2000).
sual chunk can consist of not more than one feature per dimen- Finally, it has been shown that bilateral lesions of the parietal
sion, that is, one colour, one shape primitive, and so on. A further, lobes in Balints syndrome can reduce visual processing capacity.
yet unresolved important issue in this context is the question Patients with Balints syndrome have great deficits in perceiving
whether there is also a limit in the number of possible dimensions complex visual scenes, although their ability to recognize individ-
that define a chunk. Luck and Vogel (1997) found no limit (i.e., no ual objects is usually preserved (for a review, see Rafal 1997). De-
drop in memory performance) up to a conjunction of four differ- haene and Cohen (1994) studied visual enumeration in 5 Balints
ent dimensions (colour, orientation, length, gap). However, it patients and found that these patients could enumerate sets of 1,
might be that a limitation larger than that can indeed be found. 2, and sometimes 3 items correctly, but not sets comprising more
The empirical task of the future will be to determine more pre- than 3 items. Reaction time slopes for these patients were flat for
cisely the limitations of chunking and how they relate to visual fea- set sizes of 1 and 2 items, but increased sharply for set sizes of 3
tures and dimensions. Indeed, we think that the paradigm pre- or more items. Treisman and colleagues (Friedman-Hill et al.
sented here offers a promising experimental approach to answer 1995; Robertson et al. 1997) reported another Balints patient who
questions about the nature of chunks in vision. Measuring mem- could not correctly enumerate more than one or two objects even
ory performance for a variety of stimuli and features could reveal when he was aware that more were present. In rare and extreme
the basic dimensions and features in vision in a very straightfor- cases, Balints patients report seeing only one object when pre-
ward way: If adding the feature in question to the stimuli leaves sented with multiple objects (e.g., Coslett & Saffran 1991).
the memory capacity (in terms of number of objects) unaffected, Specific visual limits or general memory/attention limits?
one may conclude that it is really a basic visual feature, forming an Cowan views such evidence as continuous with data concerning
elementary part of a visual chunk. the number of chunks which can be simultaneously active in short
Cowan defines a chunk as a collection of concepts that have term memory (STM). In contrast, we think there are good reasons
strong associations to one another and much weaker associations to resist this reinterpretation, and to view the limits on visual pro-
to other chunks currently in use. This definition does not impose cessing as separate from those involving verbal and other non vi-
any constraints on the nature and number of elements that can be sual material. (In this respect we take a position similar to that of
bound into a chunk. Our experiment is a demonstration that such Miller 1956 who suspected that STM limits and subitizing limits
limitations exist, and that their analysis may lead to important in- were independent.) Given space restrictions, we will largely re-
sights into properties of visual memory. strict our discussion of this issue to the evidence which Cowan
does discuss in his target article: subitizing (wherein observers can
determine the cardinality of sets with less than 5 items roughly in
parallel and without errors) and multiple object tracking (MOT;

wherein observers can attentionally track up to 45 indepen- can track 4 dots in a field of 8 dots, but completely fail when try-
dently and unpredictably moving identical items in a field of iden- ing to track 4 line endpoints in a field of 4 lines (and thus, 8 end-
tical distractors). points). In general, very specific rules involving connectedness
Cowan presents only a few arguments for interpreting these and part-structure seem to determine whether a feature cluster
phenomena in terms continuous with general STM limits. For can be tracked in MOT (Scholl et al., in press). Similarly, in visual
MOT he provides no arguments, simply stating that one could use short-term memory studies using a change detection paradigm,
a general STM-based theory to explain performance. (Such an ex- color and orientation features are best remembered if they belong
planation, it seems to us, could not easily account for the strong to the same part of an object and less well remembered if they be-
dependency of MOT performance on subtle visual details such as long to different parts of an object (Xu, submitted). All of these
the type of accretion and deletion behind occluders; Scholl & constraints are in marked contrast to the robustness and flexibil-
Pylyshyn 1999). For subitizing, he notes the vision-based theory ity of potential STM chunks with verbal materials.
of Trick and Pylyshyn (1994a), and argues against it mainly by ap- We think the considerations discussed here provide good rea-
peal to two phenomena. First, he suggests that the pop-out al- sons for thinking that the limits of approximately 4 involved in var-
luded to by Trick and Pylyshyn can also occur for larger numbers ious types of visual processing are distinct from other similar STM
of items, for example when all of the eggs [in a carton] pop out limits. We remain agnostic on the question of why there should ex-
against the surrounding carton (sect. 3.3.2). This, however, is ist similar independent limits. It could be for the teleological and
clearly not the type of pop-out that Trick and Pylyshyn (and oth- computational reasons discussed by Cowan (in sect, 4.1), or it
ers who have investigated visual search) have in mind, since the could be as George Miller (1956) suspected of the similarity of
eggs in this case do not pop out as individuals, but as a group. Sec- memory capacity and subitizing limitations nothing more than
ond, Cowan suggests that focused central attention is more im- a coincidence.
portant to enumeration than is suggested by Trick and Pylyshyns
theory, since other researchers (Atkinson et a1. 1976; Simon &
Vaishnavi 1996) studying the enumeration of dots in afterimages
have claimed that observers cannot enumerate sets greater than 4
without eye movements. This claim is false, however, and the lim-
its these investigators found were due to the confounding effects
How unitary is the capacity-limited
of crowding (He et al. 1997). attentional focus?
Beyond Cowans arguments, we think there are several addi-
tional reasons to view these limits as distinct from those involved Torsten Schubert and Peter A. Frensch
in verbal STM. First, viewing them as identical seems to necessi- Department of Psychology, Humboldt-University Berlin, D-10117 Berlin,
Germany. {torsten.schubert; peter.frensch}@psychologie.hu-berlin.de
tate a prediction that one should not be able to track 4 targets
in the MOT task and simultaneously acquire and hold 4 verbally-
presented items in STM. However, this is trivial to do, and such Abstract: Cowan assumes a unitary capacity-limited attentional focus. We
argue that two main problems need to be solved before this assumption
tasks seem not to interfere at all. (In an informal test, two ob- can complement theoretical knowledge about human cognition. First, it
servers tracked 4 in 8 items for 10 sec with an accuracy of 87.5% needs to be clarified what exactly the nature of the elements (chunks)
averaged over 10 trials. When they also had to remember 4 ran- within the attentional focus is. Second, an elaborated process model needs
dom digits presented auditorily as the targets were being speci- to be developed and testable assumptions about the proposed capacity
fied, they tracked with an accuracy of 92.5%, and made no errors limitation need to be formulated.
on the memory task.) Cowan notes in section 4.2 of the target ar-
ticle that such evidence against a single capacity limit could be ex- One of the main contributions of Cowans important target article
plained away by appeal to attentional switching back and forth be- is the assumption of a unitary limitation of the attentional focus.
tween the two tasks, but in this respect MOT is an ideal foil, since Cowans arguments in favor of this assumption should reinforce
one can succeed in the task only by continuous tracking (Pylyshyn the current discussion about the nature of unitary (e.g., Baddeley
& Storm 1988). Second, an explanation based on a single general 1986; Norman & Shallice 1986) or distributed attentional mecha-
limitation of memory or attention predicts that these limits should nisms (Allport 1987; Meyer & Kieras 1997; Neumann 1987). Al-
stand or fall together in neuropsychological impairments, which though we agree that this assumption is intriguing, we are some-
they do not. For example, none of the Balints patients mentioned what disappointed by its theoretical elaboration and by the
above exhibited deficits in short-term memory span. There are pa- absence of significant support advanced in its favor by Cowan.
tients who, after lesions in the left hemisphere language areas, ex- As supporting evidence for his assumption, Cowan lists differ-
hibited reduced STM span despite normal speech production in ent studies that all yield performance restrictions of about 4 items
some cases (e.g., Baddeley 1986; Shallice & Warrington 1970). in different experimental contexts and over a wide range of stim-
However, none of these patients showed any signs of Balints ulus materials, for example, dots, digits, screen locations, auditory
symptoms or deficits in visual processing. Moreover, although signals, and so forth. Because all these studies somehow yield the
these patients showed very poor retention of auditorily presented number 4, Cowans main argument is that there should be a uni-
digits, with a span in the region of two items, they usually showed tary mechanism underlying this limitation. However, for this ar-
better retention of visually presented digits, with a span in the re- gument to be convincing, it needs to be shown, first, that the items
gion of 4 or 5. These double dissociations in lesion sites and pa- across the different reported experimental contexts are compara-
tient performance argue strongly against the notion that a com- ble entities. Second, one would need to describe an elaborated
mon capacity limitation underlies capacity limited performance in process model with a set of mechanisms formulated that allows an
both verbal and visual tasks. integrated understanding of the findings.
Visual objects vs. chunks in memory. The view that these lim- Unfortunately, the present work falls somewhat short on both
itations in visual processing are distinct from those involved in of these points. For example, Cowan uses the concept of chunks
other memory phenomena is further strengthened by the fact that in stressing the equality of items in different experimental con-
the units of processing in each case are quite different. The texts. However, he does not formulate a convincing operational
chunks of memory can be almost infinitely flexible in their com- definition of what exactly a chunk is. How do we know that 4 dots,
position, and are thus defined by Cowan and others simply in screen locations, or digits correspond to 4 chunks as, for example,
terms of association networks (see sect. 1.3). This flexibility is in 4 words might? The main question is: How can we measure or de-
marked contrast to the units of visual attention visual objects fine chunks independently of the experimental context we are
which appear to be characterized by highly constrained and in- dealing with in a concrete experimental situation? If there is no
flexible rules (Scholl, in press). In MOT, for instance, observers sufficiently constraining definition, what prevents us from arguing

that 4 items held in visual short-term memory (STM) and 4 legs What about studies suggesting no dual-task costs at all (e.g.,
of a chair have a common causality in the limitation of subjects at- Hazeltine et al. 2000; Schumacher et a1. 1998)? One important
tentional focus? feature of such studies appears to be that the reduction of dual-
Furthermore, Cowan does not formulate an elaborated process task costs emerges after long, specialized training. It is thus plau-
model to explain the findings in different experimental contexts on sible that training leads to an over learned mapping of stimuli and
the basis of a common set of mechanisms. Lacking such a model, responses and consequently to automatic activation of the re-
Cowans position is descriptive but not explanatory. Thus, one sponse when a stimulus is presented. In this case, no conscious de-
could argue that the reported results concerning STM (Sperling cision is necessary to select the appropriate response for a special
1960), articulatory loop (Baddeley 1986), visual search (Fisher stimulus, and the proposed capacity limit on the attentional focus
1984), enumeration (Trick & Pylyshyn 1993), and so on, are best is not observed (see also Greenwald & Shulman 1973).
explained by different (sometimes even computational) models as- The above conjectures show that the assumption of a unitary ca-
suming separate attentional and memory mechanisms (Allport pacity limit on conscious information processing may shed new
1987; Neumann 1987). The alternative assumption of a unitary at- light on different and, even, contradictory findings in other fields,
tentional mechanism as a common cause for the reported results that have been investigated in isolation. Given this and given an
must remain highly speculative until tested. understanding of the elements of consciousness (chunks) as well
On the basis of an elaborated process model, one could formu- as an elaborated process model, the assumption of a unitary ca-
late specific predictions for an empirical test of the common-cause pacity limitation could be fruitfully included in a broader theory
hypothesis. For example, if a process or mechanism A (e.g., the at- of the human cognitive system.
tentional focus) is identified as essential for tasks 1, 2, 3, . . . , and
n, then one should be able to find a factor X, the manipulation of
which will have equal consequences on subjects performance in
tasks 1, 2, 3, . . . , and n. In contrast the manipulation of a process
B, essential only for tasks 1 and n, should influence performance Dispelling the magic: Towards memory
only in tasks 1 and n, and not in the remaining tasks. Converging without capacity
evidence of this kind would provide most valuable to support for
assumption of a unitary capacity limited attentional focus. Niels A. Taatgen
If these caveats can be resolved, the assumption could signifi- Department of Artificial Intelligence, University of Groningen,
cantly deepen our theoretical understanding of the human cogni- 9712 TS Groningen, Netherlands. niels@tcw2.ppsw.rug.nl
tive system. To illustrate, consider a different research area that www.tcw2.ppsw.rug.nl/~niels
also deals with capacity limitations, research on dual tasks. Stud-
ies using the psychological refractory period (PRP) paradigm, in Abstract: The limited capacity for unrelated things is a fact that needs to
which subjects perform two choice-reaction tasks simultaneously, be explained by a general theory of memory, rather than being itself used
as a means of explaining data. A pure storage capacity is therefore not the
have yielded rather contradictory results on capacity limitations. right assumption for memory research. Instead an explanation is needed
Many studies with this paradigm have provided support for a of how capacity limitations arise from the interaction between the envi-
capacity-limited central mechanism related to response selection ronment and the cognitive system. The ACT-R architecture, a theory with-
(McCann & Johnston 1993; Pashler 1994; Schubert 1999; Welford out working memory but a long-term memory based on activation, may
1980). These studies suggest that only one response selection can provide such an explanation.
take place at a time, thus limiting the capacity of a central atten-
tion mechanism to 1 item (chunk) contrary to Cowans proposal of The goal of science has always been to show that things in the
4 chunks. However, other contradictory findings with the PRP world that appear to be accidental can be explained by a set of sys-
paradigm suggest no capacity limitation at response selection at all tematic and fundamental principles. Millers (1956) magical num-
(Hazeltine et al. 2000; Schumacher et al. 1998). The latter results ber seven and subsequent theories based on the idea are attempts
stimulated Meyer and Kieras (1997) to propose a scheduling ac- to find such principles. Cowan tells us that the magical number is
count (EPIC) for dual-task situations in particular and sensori- not seven, but actually four. Still the word magical lingers
motor tasks, as well as working memory tasks in general. Accord- around this mysterious capacity. My proposal is not to attack the
ing to EPIC, elements of the task to be performed are maintained seven aspect of the principle, but rather the magical part of it,
in a working memory system without any limiting attentional ca- since we all know that magic doesnt really exist. Whereas in
pacity. Executive processes allow the scheduling of task processes Millers original article short-term capacity was just an empirical
according to subjects specific instructions and goals. Cowan ex- fact, it has subsequently grown into a theory that people actually
presses doubts that this scheduling theory can account for the have a pure storage capacity.
findings on attentional and memory limitations. However, looking Let us elaborate on this idea. If we take capacity seriously, the
at the contradictory evidence and theoretical positions, the ques- number of items that can be stored by an individual has to be an
tion arises whether there are different task-specific working mem- integer. A capacity of 3.5 only makes sense as a group average, not
ories with different capacity limitations, for example, one for dual as a property of an individual. An individual can either retain three
tasks and one for storage tasks, or whether there exists no central items or four items, not three-and-a-half. An individual capacity
capacity limitation at all. of 3.5 only has meaning if the individual can sometimes remem-
We assume that the assumption of a capacity-limited atten- ber three items, and sometimes four items. But this is hard to rec-
tional focus may account for the contradictory results in dual-task oncile with the idea of a fixed capacity. It becomes even harder to
research if one localizes the capacity limitation strictly at the level explain development. Even according to Cowans own data, the
of conscious information processing. A plausible explanation is capacity of adults is larger than the capacity of children. But how
that in dual-task studies a capacity limitation of the central mech- then does this capacity grow? Are there sudden increases in which
anism can be observed only when subjects carry out the tasks in a the capacity is incremented by one?
relatively unlearned state. In this case, subjects probably make a The problem with the target article is that it already assumes
conscious decision which response to map to which stimulus when there is a capacity limit, and that it can be studied separately from
performing the task. The conscious decision requires maintaining the rest of memory. If one wants to prove there is indeed a limit-
a set of different stimuli and different responses in working mem- capacity short-term store, the relation to long-term memory
ory together with a set of S-R mapping rules. One could easily (LTM) has to be taken into account. When something drops out
imagine that, in this case, the number of chunks in working mem- of short term memory (STM), is it really gone? Sometimes the ex-
ory exceeds the proposed limit of 4, thus causing a deterioration act information is irretrievable, but the vast literature on implicit
of dual-task processing. learning and priming suggests that everything that happens in

STM has some long-term impact. So what of the alternative ac- high capacity individual. The individual differences were pro-
count, that STM is no separate entity, but just a part of LTM? This duced by variation of an ACT-R parameter that controls the
would be a much more parsimonious solution, provided it can ex- spread of activation (based on Lovett et al. 1997). The reason why
plain the empirical facts of a limited short-term store. the curve drops off so dramatically at around four items has noth-
I would like to argue that an explanation of short-term store ing to do with the number four itself. It rather has do to with the
based on properties of LTM is much more interesting than as- fact that as the string of numbers grows, the effects of decay, in-
suming a separate entity. Why is the capacity four, and not five? A terference, and the increased probability of doing something
theory that proposes a buffer of limited size does not provide any wrong if more responses are required are multiplied, and cause
answers. Take for example the subitizing phenomenon, the fact performance to drop suddenly at this point.
that people seem to be able to recognize up to four dots in the vi-
sual field, but have to count if there are more. One could postu-
late the theory that the visual system has an built-in capacity to
recognize up to four things, and be done. Peterson and Simon
(2000) offer an alternative account. According to their theory, the How to interface cognitive psychology
visual system can immediately recognize a set of dots, if it has seen
these dots in the same array before often enough. The number of with cognitive neuroscience?
possible configurations of dots increases exponentially by the
Hannu Tiitinen
number of dots. Therefore, the human visual system receives
Department of Psychology, University of Helsinki, Helsinki FIH-00014,
enough examples of four-dot configurations to recognize any of
Finland. hannu.tiitinen@helsinki.fi www.helsinki.fi/hum/ylpsy
them instantly, but not of five or more. Except of course when a
particular configuration occurs often enough: anyone can recog-
Abstract: Cowans analysis of human short-term memory (STM) and at-
nize the five-dot pattern on a die instantly. The advantage of the tention in terms of processing limits in the range of 4 items (or chunks)
Peterson and Simon account is that they show how the seemingly is discussed from the point of view of cognitive neuroscience. Although,
magical number four can be explained by an interaction between Cowan already provides many important theoretical insights, we need to
environment and the cognitive system. learn more about how to build further bridges between cognitive psy-
Short-term memory capacity is not something that can be used chology and cognitive neuroscience.
to explain the outcomes of experiments, but is rather something
that needs to be explained itself. One possible explanation is the Cowans target article on the limits of mental storage capacity de-
one offered by the ACT-R architecture (Anderson & Lebiere lineates experimental observations on short-term memory (STM)
1998). ACT-R has a long-term declarative memory that also serves into either pure or compound limits with an average of 4 chunks
as working memory. To keep track of the current context, a single- and equates this limit with human attentional mechanisms. In my
item focus of attention is used. All items that have to be memo- view, this is exactly what is needed in cognitive psychology and
rized are stored in declarative memory. Since declarative memory even more so in cognitive neuroscience today.
is activation-based, interference and decay can produce the same In cognitive psychology, the quest for an understanding of hu-
sort of effects usually assumed to be produced by limited STM. man attentional processes has its roots in the groundbreaking
These limitations are, however, context dependent: if there are as- work of Cherry (1953), Miller (1956), and Broadbent (1958). As
sociations between the items to be memorized or with other items Cowan points out, sometimes these major findings become such
in memory, it is easier to retrieve the information. Short-term a major influence, far beyond the expectations of individual re-
memory without context is only important if one presumes its ca- searchers, that they can distort scientific progress. Over the past
pacity is a fundamental property. Short-term memory within a decades, observations have accumulated at such a rapid rate that
context is much more useful. I have demonstrated (Taatgen 1999a; it is very important to evaluate related findings and try to find
1999b) that individual differences on simple memory task might some common denominators in the data. Cowans target article is
be used to explain individual differences in skill acquisition. an important theoretical organizer and a logical continuation of his
Figure 1 shows some results from an ACT-R model of short- previous work (1988; 1995).
term store that I have adapted from an earlier version (Taatgen Current brain research has a tendency to focus on the study of
1999a). The original model memorized a list of up to ten digits, a given response and its dynamics, obtained with a given method,
and attempted to reproduce them. It neatly reproduced Millers but it is the theory, not the method, that should guide the re-
7 6 2 effect. Since the original model was allowed to rehearse, I searcher. This brings us to a few specific comments on the target
removed the rehearsal, and obtained the results in Figure 1: the article.
magical number four, but without any internal capacity limita- First, I would like to question the assumption of human brain
tions. The figure shows three curves, a simulated low, average and oscillations, especially in the gamma (ca. 40 Hz) range, serving as
a binding mechanism for sensory input and, in the present con-
text, as the neuronal mechanism underlying chunking of informa-
tion. The role of gamma activity is still under debate and, although
the linking of STM and chunking with oscillatory brain activity is
an attractive new idea, one must be very cautious in interpreting
the data, such as those of Lisman and Idiart (1995), or as suggested
by Cowan (sect. 4.1.2). As Cowan shows, one can calculate and
show correlations in cycles and subcycles, and estimate the ra-
tio of slow and fast oscillatory rhythms in various ways, ad infini-
tum. Furthermore, 40-Hz oscillations are actually much more
widespread in the frequency domain than is generally assumed
and highly susceptible to, for example, the use of recording and
filtering settings. Is this line of correlative research fruitful and
promising enough? More specifically, can Cowan envision other
prospective avenues of research which might help bridge the the-
oretical concepts of cognitive psychology and the measures of cog-
nitive neuroscience?
In bridging the gap between these two domains in the research
Figure 1 (Taatgen). of STM and attention, at least two issues suggest themselves for

further inspection. (1) Experiments explicitly designed within the pacity limits. The essence of this statement is substantiated in
theoretical framework provided by Cowan and (2) the inherent other parts of his paper, and I admit that I fully agree with it. Nev-
limitations in brain research techniques. ertheless, I would like to submit a proposal which is guided by a
Most of the neurophysiological studies cited in the target arti- comparative perspective, suggesting that we should extend the re-
cle do not seem to specifically address STM. For example, Gray search focus from the STM mechanisms of human beings to those
et al. (1989) studied cats using invasive measures, and Tiitinen et of animals.
al. (1993), using EEG, focused only on the attentional effects re- Several studies on animals have shown that limitations in the ca-
flected in the human transient 40-Hz response. The latter obser- pacity to process acquired information are not unique to humans,
vations were extended by Cowan (sect. 4.1.4) as consistent with but found in nonverbal organisms as well (Chen et al.; Koehler
the idea of a STM storage capacity of 4 chunks when taken to- 1954; Roitblat et al. 1991; Terrace 1987; 1998; Todt et al. 2000). It
gether with Cowan et al.s (1999; Fig. 4) very interesting simple may therefore be wise to select animal models that either allow in-
relationship between attended and unattended speech, with at- quiries into STM mechanisms which are difficult to address in hu-
tentional allocation presumably drawn from the same processes in mans, or that permit us to uncover memory properties which are
both conditions. biologically basic and thus point to precursors of accomplishments
This interpretation is however complicated by the fact that Ti- regarded as human characteristics. A paradigm for such an ap-
itinen et al. (1993), using sinusoidal stimulation only, observed not proach was recently published by Hultsch (1992; 1993). She stud-
only the attentional enhancement in the 40-Hz range, but a promi- ied acquisition of serially structured information in the nightin-
nent response under passive (reading) conditions too. This already gale (Luscinia megarhynchos), a songbird who is able to learn and
poses two variables that need to be taken into account in the quest memorise a vocal repertoire of more than 200 different sound pat-
for pure STM limits: the type of stimulation used (sinusoids vs. terns (5 songs). In the following, I first recapitulate some back-
speech) and brain activity related solely to attention versus that ob- ground facts, then outline the methods and results of Hultschs ex-
served in passive conditions (such as reading). Sensory-specific, periments, and finally discuss implications for special issues
task-independent brain processes (easily observable in passive con- treated in the target article.
ditions) might actually have a much more crucial role in higher- When singing, most songbirds produce a series of well-struc-
order stages of information processing than previously assumed tured vocal patterns that typically have a duration of a few seconds
(Tiitinen et al. 1994; Tiitinen & May, preprint). This, I feel, must and are called songs (Fig. 1). Although each single song encodes
be emphasized despite the fact that our understanding of sensory- information about several biological details, for example, both in-
level memory processes is still evolving (see, e.g., May et al. 1999). dividual and species-specific cues, many species develop and use
A more straightforward way to understand the STM limitation repertoires of several different types of songs (Catchpole & Slater
might be the design of experiments based on the framework pro- 1995; Kroodsma & Miller 1996). Song development is a matter of
vided by Cowan. This impressive target article already suggests vocal learning (Hultsch 1991; Marler 1976; 1991; Nottebohm
several interesting research avenues for brain research: For ex- 1993). This accomplishment covaries, however, with the hierachy
ample, the results of Cowan et al. (1999) could be further ex- level of learning stimuli: Whereas birds normally copy the acousti-
tended in the context of EEG and MEG measurements, in which cal patterns of single songs, they appear less precise when learn-
one could also take into account the problems mentioned above. ing at a higher level of song organisation. That is, their acquisition
An equally interesting issue is that of the flexible use of chunk of information encoded in a series of different songs can be con-
sizes, which can range from small groups to supergroups (Erics- strained by a mechanism called package formation (Hultsch et
son et al. 1980; Ericsson 1985) or active superconcepts (Shastri al. 1995).
& Ajjanagadde 1993). These not only provide humans with a pow- The phenomenon of package formation reflects properties of
erful operating advantage in complex cognitive environments, but STM involved in the acquisition of songs and indicates that this
might even help explain the observed individual differences in achievement is mediated by a process reminiscent of the chunk-
STM limits. These observations can readily be made in fMRI and/ ing of items in human serial item learning. The phenomenon was
or PET measurement, which might then shed light on where in discovered by analyses of groups of sequentially associated songs,
the brain the STM chunks of varying size are located. These, and or song packages, respectively. These were developed by night-
similar attempts, should eventually provide us with a map of the ingales, who during their first weeks of life had been exposed to
architecture of cognition and one of its important aspects: flex- long strings of stimulus songs. Such packages had a mean size of
ible memory storage and attentional influences. 4 6 2 songs (Hultsch & Todt 1989). After several studies on this
matter, it became evident that the packages were not a result of
song recall or memory retrieval, but induced by STM properties
upon early auditory exposure to learning stimuli. In addition,
package formation is explained by a model postulating a joint
Studies of STM properties in animals may operation of the STM and a battery of submemories: When a
help us better understand the nature young bird is exposed to a succession of several new songs, his STM
processes information of about 26 different songs, and then
of our own storage limitations: transmits the information to a specific submemory where the sec-
The case of birdsong acquisition ond step, that is, a longterm storage of song material, takes place.
It was hypothesised that, only after the transmission is the STM
Dietmar Todt ready to process a further number of songs, which are then stored
Institute of Biology: Behavioural Biology, FU Berlin, D-12163 Berlin, in a different submemory (Todt & Hultsch 1996; 1998). In other
Germany. todt@zedat.fu-berlin.de
words, the model predicts that the STM segments a given se-
www.verhaltensbiologie.fu-berlin.de
quence of learning stimuli and the battery of submemories con-
solidates this effect by processing the segments as different pack-
Abstract: I like Cowans review of STM properties and especially his sug-
gestions on the role of attention. I missed, however, a consideration of
ages of songs.
studies which provide evidence for STM properties in animals. In my com- To further test the memory model and, in particular, to clarify
mentary, I argue that such evidence can elucidate the biological basis of whether the segmentation would be a genuinely unit related (i.e.,
storage limitations, validating this view by discussing mechanisms which information constrained) process or whether time related effects
constrain the acquisition of serial information in songbirds. could also play a role, Hultsch examined how young nightingales
would cope with learning programs that differed in the rate of
In the introduction to his target article (sect. 1, para. 2) Cowan em- stimulus songs (Hultsch 1992). Programs were prepared by mod-
phasises that we are still uncertain as to the nature of storage ca- ifying the duration of silent intervals between successive stimulus

Figure 1 (Todt). Frequency spectrograms of three nightingale songs (section of a longer singing episode). These songs are developed
by vocal imitation, and during their early acquisition processed analogous to chunks (see text).
songs. In one test series, the normal intervals (duration: 4 sec) The good accord between properties of song acquisition in
were shortened to just 1 sec; thus 1 min included 13 learning stim- birds and properties of STM in humans, prompts three final com-
uli here (dense designs). In a second test series, such intervals ments. First, it seems evident that a study of STM properties in
were prolonged up to 10 sec; thus 1 min included only 5 learning animals can help to elucidate the nature of our own storage limi-
stimuli here (spaced design). For a contro1, the subjects were tations (Roitblat et al. 1991). Second, song acquisition is a useful
also exposed to temporally unmodified series (norma1 design biological model for comparative studies, for example, one can
that is, 8 learning stimuli per min). In order to adjust the experi- test memory mechanisms by applying learning stimuli that are
mental design to the methods of former studies, each of the three completely new for a young subject and do not produce cognitive
stimulus sequences was composed of 20 different types of songs problems as do so many learning stimuli in human serial item
which the birds could not experience in any other sequence or learning (Todt et al. 2000). Third, the finding that birds use mem-
learning program. Analyses of singing behaviours performed by ory mechanisms which operate similarly to ours points to inter-
the trained birds yielded results that can make an interesting con- esting parallels in the evolution. Since the brains of birds are not
tribution to the issues raised by the target article (Fig. 2). homologous to the those of mammals (Konishi 1989; Pepperberg
Above all, the study showed that the early segmentation song 1993), similar operational properties suggest that they have evolved
stimulus sequences is controlled by two components: a unit (or in- because they allow an optimal solution to a given problem; that is,
formation) based capacity buffer (evidence: constraints uncov- here, a strategic acquisition of perceptional information.
ered by the dense program) and a time window based gating
mechanism (evidence: constraints uncovered by the spaced pro- ACKNOWLEDGMENT
gram). The capacity buffer limited the sizes of stored packages at The memory research on nightingales is supported by DFG grants.
3 5 songs. The time related component, on the other hand, lim-
ited such sizes by a time window of ca. 32 sec. Based on new evi-
dence (Hultsch et al. 1999). This window indicates a specific span
of attention as suggested in Cowans article.
Memory limits: Give us an answer!
John N. Towse
University of London, Egham, Surrey, TW20 OEX, United Kingdom.
j.towse@rhbnc.ac.uk www.pc.rhbnc.ac.uk/jt/jt.html
Abstract: Cowan has written a meticulous and thought-provoking review

of the literature on short-term memory. However, reflections on one area
of evidence, that of working memory span, shows the extent to which the
research debate can be circumscribed by choice of experimental para-
digms.
In the entertaining radio series, The hitch-hikers guide to the

galaxy, Douglas Adams recounts how, many millions of years ago,
a race of hyper-intelligent, pan-dimensional beings became so an-
noyed at all the constant bickering over the meaning of life, that
they decided to solve the arguments once and for all (Adams
1985). They created Deep Thought, the most awesome computer
imaginable and then gave him the task of finding the Ultimate An-
swer To Life, The Universe, and Everything. Deep Thought took
his time to consider this challenge (seven and a half million years),
but in the end he was ready. To the hushed anticipation of the uni-
verse, Deep Thought reluctantly announced that The Answer was
Figure 2 (Todt). Illustration of package sizes (number of associ- in fact 42.
ated song types) that nightingales (n58) developed under three This produced understandable commotion. It was hardly a sat-
different learning programs. Filled circles show: Packages devel- isfactory pronouncement, a judgment that would allow everyone
oped from the dense program (left) were not larger than those to sleep easy at night. But as Deep Thought pointed out, I think
developed from the normal program (middle). This suggested the problem such as it was was too broadly based. You never ac-
unit-related capacity constraints on data processing to account for tually stated what the question was. He had found the answer, but
package formation. On the other hand, packages developed from what exactly did the ultimate question ask? Deep Thought
the spaced program (right) were significantly smaller (see open couldnt say, and therefore he had to design an even more power-
circle) than predicted by a genuinely unit-controlled limitation ful computer which could formulate the question to the ultimate
process. This suggested that the postulated segmentation was con- answer (sadly, this computer, Earth, was destroyed just before it
currently controlled by time factors (see Hultsch 1992). had completed the job, but thats another story).

The relevance of this whimsy lies in the illustration of how we Neural mechanism for the magical number 4:
can lose sight of the question that were asking when searching for Competitive interactions and nonlinear
an answer. Cowan has amassed an impressive array of data that oscillation
leads him to believe that The Ultimate Answer to Memory is 4.
By constraining the circumstances for admissible evidence bear- Marius Usher,a Jonathan D. Cohen,b Henk Haarmann,c
ing on this question, Cowan has refined Millers earlier conjecture and David Hornd
that the answer is 7. In fact, whether you prefer 4 or 7 or a
School of Psychology, Birkbeck College, University of London, London
some other number probably relates to whether you feel the var- WCJE 7HX, United Kingdom; bDepartment of Psychology, Princeton
ious measurement conditions are appropriate or not, and Cowans University, Princeton, NJ 08544; cDepartment of Hearing and Speech
line of reasoning certainly needs to be taken seriously. Sciences, University of Maryland, College Park, MD 20742; dDepartment of
In the case of short-term memory, though, surely the question Physics, Tel Aviv University, Tel Aviv, Israel. m.usher@bbk.ac.uk
is quite transparent What is the limit of mental storage capac- jdc@princeton.edu hhaarmann@hesp.umd.edu
ity? So wheres the fuss? Well, although we can ask this question horn@neuron.tau.ac.il www.psyc.bbk.ac.uk/staff/homepage/mu.html
and find an answer (even perhaps find a coherent answer) does www.bsos.umd.edu/hesp/haarmann.html
that help us identify whether it is the best or only question to be www.neuron.tau.ac.il/~horn
asking? Unfortunately, this is less clear.
This problem is hardly Cowans fault. Consider, for example, the Abstract: The aim of our commentary is to strengthen Cowans proposal
area of working memory capacity, covered in the target article. for an inherent capacity limitation in STM by suggesting a neurobiologi-
cal mechanism based on competitive networks and nonlinear oscillations
Working memory capacity is conventionally estimated via working that avoids some of the shortcomings of the scheme discussed in the tar-
memory span tasks. These are supposed to require the simultane- get article (Lisman & Idiart 1995).
ous combination of processing and retention. They exist in
several forms such as reading span where individuals read a series Cowan interprets the capacity limitation of 4 6 1 items, in rela-
of unrelated sentences for comprehension and afterwards attempt tion to a bottleneck on the focus of attention. While this may be a
to recall each sentence-terminal word (Daneman & Carpenter plausible explanation, it is somehow paradoxical that recall of un-
1980). Reading span is calculated on the basis of the number of attended material should provide a measure for the capacity of the
sentences that can be presented and followed by correct recall of focus of attention; with diverted attention, recall may be mediated
the relevant words. Now, one can argue about one or other theo- by the residual decay limited components in STM. Nevertheless,
retical account of performance (e.g., Towse & Houston-Price, in we believe that the rest of the reviewed data provides convincing
press) and doing so can certainly lead to a richer appreciation of support for a capacity limitation of about 4 items in STM. This
working memory. It is important to recognise, however, that these raises the challenge of providing a plausible neurobiological mech-
theoretical debates all centre around what happens to span scores, anism for it. Cowan adapted the model of Lisman and Idiart
that is, the maximum number of things that can be remembered. (1995), explaining the capacity limitation in terms of two wave fre-
Under such circumstances, therefore, there are few options avail- quencies. While this adaptation needs to be tested, its likelihood
able to the scholar other than to do what Cowan has and derive an for confirmation is low, as the g and u waves used in the original
average value for memory capacity. schemes were based on empirical estimates of neurophysiological
So although the target article briefly considers alternatives to data. Moreover the frequency of these oscillations is broad and
the storage capacity account and finds them wanting (sect. 4.3), therefore their ratio is likely to fluctuate widely rather than pro-
the evidential basis on which to evaluate the different accounts vide a stable value corresponding to a specific capacity.
simply is not balanced. With so many studies of short-term mem- An alternative neurobiological account does not rely on precise
ory, let alone working memory, based on span scores or their frequency values but explains the capacity limitation in terms of
equivalent, we run the danger of rigging the contest. That is, pre- inherent properties of competitive networks (Horn & Usher 1991;
determining the sort of answer we will look for and therefore find. 1992; Usher & Cohen 1999). The main idea for consideration is
Recent, as yet unpublished collaborative work with Graham that while LTM is mediated by structural changes in connectivity,
Hitch, Una Hutton, and Zo Hamilton has begun to explore other STM (which is associated with awareness) is mediated by neuro-
ways of asking questions about working memory. Rather than fo- electric reverberations (Hebb 1949), subject to competitive inter-
cus on ( just) the number of items that children can remember, we actions, whose need has often been discussed in experimental and
have asked whether, in variants of working memory span tests, computational neuroscience (Desimone 1998; Grossberg I976;
children show meaningful and reliable variations in the temporal Usher & Cohen 1999; Usher & McClelland 1995) in relation to at-
endurance of their memory traces. Potentially at least, some cog- tentional selection. Typically selection is implemented in compet-
nitive skills may rely more on the extent to which representations itive models by a mechanism of strong lateral inhibition, resulting
can be preserved, as opposed to the number of stimulus neigh- in a winner-take-all system. We have recently presented a model
bours that can be tolerated in memory tests. Early indications of which proposed that the lateral inhibition can be modulated in re-
this research programme seem promising, but in the present con- lation to task demand (Usher & Cohen 1999). The model is de-
text, it is perhaps the attempt to seek alternative dependent vari- scribed by the following equation:
ables that is most relevant. Even if endurance measures turn out
to be redundant, or ineffective, that itself seems an important con- dxi /dt 5 2xi 1 a F(xi ) 2 bSj? i F(xj ) 1 Ii 1 noise
clusion to reach because thus far one must assume span gives all
the information one needs. Strong faith in measures of memory where xi is the activation of the i representation, F(x) 5 x/(11 x)
size alone may permit us to find The Answer, but at what cost to is the activation-functions (see e.g., Tsodyks et al. 1998), Ii is the
a full understanding of immediate memory? sensory input to each representation, a corresponds to recurrent
self excitation and b to lateral inhibition. The self excitation al-
ACKNOWLEDGMENTS lows the maintenance of activation after the input Ii is turned off.
The commentary is underpinned by support from the ESRC, grant The inhibition parameter is set high when selection is required
R000222789 (awarded to John Towse, Graham Hitch, and Una Hutton). and moderate when multiple items need to be maintained to-
Discussions with Robin Walker are also appreciated. gether, as in immediate recall tasks. In this model we showed that
moderate levels of lateral inhibition allow the coactivation of sev-
eral memory representation. More interestingly, the system
shows a sharp capacity limitation where only a small number of
items can be simultaneously maintained. This capacity depends
on the excitation and inhibition parameters and is within the

range of 3 5 items when these parameters are chosen as a 5 2 in a distinct phase. This characteristic does not depend on the pre-
and .15 , b , .20. cise values of parameters (except that they need to be within the
Functional considerations demonstrate the system cannot in- range leading to staggered oscillations) and is a result of the fact
crease its capacity beyond this range by diminishing the inhibition that it is not possible to compress more than 5 narrow non-linear
b parameter even further. Due to the recurrent excitation and to oscillations (one for each activated representation) within each cy-
small overlaps in the input (or in the representations themselves) cle of the whole system. This mechanism explains the 4 6 1
a minimal amount of lateral inhibition is required to prevent the capacity limitation, however, unlike the model described in the
activation to spread to all the memory representations. Assume, target article (Lisman & Idiart 1995). It is robust to parameter
for example, that when the representation i receives input, Ii is rel- changes and can function for a wide range of frequencies, result-
atively high (.33 in the simulations reported), while other repre- ing from competitive interactions between nonlinear oscillators.
sentations j ? i receive a much diminished input (Ij 5 .05, due to
overlaps). In this situation, we find that when the inhibition is
small (b , .19) activity spreads to nonsensory-activated units,
even when only a single unit, i, receives input. This unreliable re- Over the top: Are there exceptions
lation to its inputs would make the system useless. Increasing the
inhibition to b 5 .19 prevents this problem, but limits the num- to the basic capacity limit?
ber of co-active units to 3 4 items. Changes in the excitation pa-
John Wilding
rameter do not affect those considerations. If a is increased alone,
Department of Psychology, Royal Holloway, University of London, Egham
the capacity increases, however, increasing the excitation requires
Hill, Egham, Surrey TW20 0EX, United Kingdom. j.wilding@rhbnc.ac.uk
a simultaneous increase in the minimal inhibition to prevent the
unbounded spread of activation within the system. When the two
Abstract: Can we identify individuals with a larger basic capacity than
parameters are changed together to preserve reliability the ca- Cowans proposed limit? Thompson et al. (1993) claimed that Rajan Ma-
pacity remains in the 3 4 items range. hadevan had a basic memory span of 13 15 items. Some of their support-
Does then the capacity result from a specific layout of biologi- ing evidence is reconsidered and additional data are presented from study
cal parameters that characterize the system or is there a magical of another memory expert. More detailed analysis of performance in such
number 4 to be found? Two tentative answers can be suggested. cases may yield different conclusions.
In the model described so far, the capacity can be increased only
by reducing the representational overlap. Such reductions may be, Cowan argues for a capacity limit on short term memory of about
however, bounded due to another trade-off (Cohen 1996). Rep- four items (possibly rising to a maximum of six); this raises the
resentational overlap is essential for computations that perform question of whether some individuals have still larger capacity.
generalization. Thus it is possible that the system evolved so as to Thompson et al. (1993) studied the memory ability of Rajan Ma-
optimize the functions of active maintenance and generalization, hadevan. His digit span before he began deliberate practice at
providing another teleological motivation for the STM capacity memory tasks was recorded as 15 items and other evidence sup-
(cf. Kareev et al., in the target article). ported this estimate.
Another answer which explains a magical number 4 is suggested Thompson at al. (1993) argued that the lag between the end of
if one assumes that maintenance in STM requires not only the si- list presentation and response initiation should remain constant
multaneous activation of representations, but also their segmenta- until list length exceeds the basic span and rehearsal becomes nec-
tion from each other. This can be performed by a system based on essary. In support of this, they found that individuals with normal
staggered oscillations as described here, but with higher inhibition, digit span showed a constant lag until list length exceeded 7 items;
and with the addition of an adaptation/recovery variable that lag then rose steadily as list length increased. Rajan, however,
makes each representation oscillate (Horn & Usher 1991; 1992). showed no rise until list length exceeded 13 items. These data ex-
Due to the lateral inhibition one can generate a situation where emplify one of Cowans acceptable methods for measuring limited
each representation is activated in a different phase of the tempo- capacity; the discontinuity in lag is assumed to reflect the point at
ral oscillation, as illustrated in the Figure 1 (showing segmentation which basic span is exceeded. However, the data indicate a dis-
with 3 units oscillating out of a background of non-active units). continuity at seven items in normal memorizers, rather than at
Computational and mathematical analysis of this system dem- Cowans proposed limit of four items. Rajans basic span of 13
onstrated a fixed capacity of 3 5 items in the ability of this system items suggests that some individuals may have much greater ca-
to perform segmentation (Horn & Opher 1996). The system has pacity. With colleagues, I have attempted to identify other such in-
maximal stability for 3 out-of-phase oscillations. For more than 5 dividuals, using a computerised version of the task and recording
oscillations the system is unable to keep each of these oscillations additional measures, which clarify whether lag is a valid indicator
of basic span.
Lists of digits were presented visually at 2 items per sec; an as-
terisk indicated the end of the list and requested recall, which was
achieved by typing numbers on the keyboard. Time to initiate re-
call and to input each item was measured, plus the positions of er-
rors. Initially a five-item list was presented and 1ist length in-
creased by one item following a correct response and decreased
by one item following an error, matching Thompson et al.s (1993)
ascending method. Twenty trials formed a run. Data have been
collected from two memory experts, but results are discussed here
only from MB, who comp1eted 50 runs on the task. MB held the
world record for reciting the expansion of pi in 1977 and later
achieved correct recall to 7,769 places. He can recite the first
1,000 digits of pi in 170 sec. He claims special memory ability only
for numbers and uses none of the standard mnemonics, relying
only on his knowledge of mathematics, which he teaches at sec-
ondary school level. If his digit span depends on chunking based
on mathematical knowledge, we would expect a discontinuities in
Figure 1 (Usher et al.). Segmentation of three units oscillating the lag measure around the normal span and between chunks in
out of a background of non-active units. the inter-item recall latencies.

Table 1. (Wilding). Time to initiate recall(s) for lists of 5 to 17 items in a visual digit span task
List
Length 5 6 7 8 9 10 11 12 13 14 15 16 17
MP 1.1 1.3 1.3 1.3 1.3 1.6 1.5 1.5 2.1 2.1 2.3 2.1 2.6
Rajan 0.2 0.3 0.2 0.3 0.7 0.7 1.4 4.2 3.0 8.4 11.2
The mean of the longest span achieved on each run was 14.96 though he does not have a deliberately practised strategy for mem-
(range 1218, s.d. 1.4), less outstanding than Rajan, but about orisation, he tried to find some mathematical pattern or relation,
twice the norm. Table 1 gives the time to initiate recall at each list normally in each group of four.
length (successful recalls only) for MP and corresponding figures While any conclusions clearly need data from other individuals
for Rajan (estimated from Thompson et al., Fig. 2.1; absolute before they can be firmly established, these results imply that: (1)
times are not comparable, due to the different methods of record- Thompson et al.s (1993) response lag measure is not decisive as
ing used). While Rajan certainly shows a steep rise in times for lists an indication of basic span. Though no clear discontinuity in MPs
of more than 13 items, there is also some suggestion of an increase data indicated a basic span limit, discontinuities in inter-item la-
at 11 items; no data are given for lists of five or six items and pre- tencies suggested chunking was occurring. (2) items were in gen-
sumably measurement of times by stopwatch was imprecise for eral grouped into fours, but such grouping was not mandatory and
short intervals. MP, however, shows a gradual increase, with pos- up to seven items were treated as a group in some circumstances.
sible stops at 10, 13, and 17 items. There is, however, no clear dis- Such flexibility may explain the long running argument as to
continuity, which would unambiguously indicate a basic span lim- whether the limit is four or seven items and raises further ques-
itation. The other expert we tested, who overtly converted tions about the mechanisms involved.
numbers into images, showed a sharp discontinuity at 9 items, fol-
lowed by a continuous rise in times, like Thompson et al.s (1993)
subject GN, who also developed a mnemonic method, so the pat-
tern shown by MP was not some artefact of the measurement
method. Attention is not unitary
Inter-item latencies for successful recall, however, provided un-
ambiguous evidence for chunking in MPs recall; however, the pat- Geoffrey F. Woodman, Edward K. Vogel,
tern obtained was only partly consistent with Cowans argument. and Steven J. Luck
Figure 1 gives cumulative recall times for lists of 5 to 16 items. Department of Psychology, University of Iowa, Iowa City, IA 52242-1407.
Once lists exceeded 7 items, a pause occurred after the fourth {geoff-woodman; steven-luck}@uiowa.edu vogel@sdepl.ucsd.edu
www.psychology.uiowa.edu/Faculty/{woodman/woodman; luck/
item, once lists exceeded 11 items, another pause occurred after
luck}.html
the eighth item, and once lists exceeded 15 items, a third pause
occurred after the twelfth item. MP apparently divided the lists
Abstract: A primary proposal of the Cowan target article is that capacity
initially into groups of four items (though lists of 14 items seem to limits arise in working memory because only 4 chunks of information can
be divided 4 4 3 3), but the final group could be as large as be attended at one time. This implies a single, unitary attentional focus or
seven items before it was further subdivided. MP confirmed that, resource; we instead propose that relatively independent attentional mech-
anisms operate within different cognitive subsystems depending on the
demands of the current stimuli and tasks.
Cowans model of working memory posits that the limited capac-

ity of working memory is entirely caused by the limited capacity
of attention. He further proposes that a unitary mechanism of
attention operates across cognitive subsystems, with a single ca-
pacity that must be shared across subsystems. Here, we argue in
favor of an alternative view in which the brain makes use of a col-
lection of loosely interconnected attention mechanisms that op-
erate in different cognitive subsystems and that reflect the prop-
erties of the individual subsystems. We will focus specifically on
the operation of attention in three coarsely defined cognitive sub-
systems, namely perceptual analysis, working memory, and re-
sponse selection.
We would expect that different mechanisms of attention would
operate in visual perception and in visual working memory for the
simple reason that these subsystems appear to operate at very dif-
ferent rates. For example, Potter (1976) showed that observers
can identify pictures of complex real-world scenes at rates of up
to 8 pictures per second, but much slower presentation rates were
required for observers to store the scenes in working memory.
More recent studies have used a combination of psychophysical
and electrophysiological measures to demonstrate that attention
shifts at different rates in visual perception and in visual working
memory. Specifically, studies of the attentional blink phenome-
Figure 1 (Wilding). Mean cumulative recall times by MP for non, which reflects the operation of attention in working memory,
each item in lists of 5 to 16 items. Numbers on the left indicate list indicate that hundreds of milliseconds are required to shift work-
length for each function. As list length increases, each function has ing memory-level attention from one stimulus to another (Vogel
been displaced upward by 1 sec to avoid overlap. et al. 1998). In contrast, studies of visual search, which empha-

sized the operation of attention in perception, have indicated that

shifts of perceptual-level attention can occur every 50 100 msec
Authors Response
(Woodman & Luck 1999). Horowitz and Wolfe (1998) and col-
leagues have also provided evidence that the operation of atten-
tion in perception is relatively independent of memory.
Given that attention operates at different speeds in perception Metatheory of storage capacity limits
and in working memory, it would be sensible if these mechanisms
of attention operated asynchronously; otherwise, the fast percep- Nelson Cowan
tual system would be continually waiting for the slow working Department of Psychology, University of Missouri, Columbia, MO 65211.
memory system to catch up. We have recently provided evidence cowann@missouri.edu www.missouri.edu/~psycowan
for asynchronous operation of perception and working memory Abstract: Commentators expressed a wide variety of views on
by showing that shifts of attention during visual search are not whether there is a basic capacity limit of 3 to 5 chunks and, among
slowed when visual working memory is full (Woodman et al., in those who believe in it, about why it occurs. In this response, I con-
press). In this study, participants were required to maintain a set clude that the capacity limit is real and that the concept is
of object representations in visual working memory while per- strengthened by additional evidence offered by a number of com-
forming a visual search task. We observed that subjects could per- mentators. I consider various arguments why the limit occurs and
form a difficult visual search task just as efficiently when working try to organize these arguments into a conceptual framework or
memory was full as when working memory was empty, indicating metatheory of storage capacity limits meant to be useful in fu-
that perceptual-level attention can be allocated and shifted very ture research to settle the issue. I suggest that principles of mem-
efficiently even when working memory-level attention is operat- ory representation determine what parts of the representation will
ing at maximal capacity. be most prominent but that limits of attention (or of a memory
It is also reasonable to suppose that visual perception and visual store that includes only items that have been most recently at-
working memory might differ in their spatial properties as well as tended) determine the 3- to 5-chunk capacity limit.
in their temporal properties. In particular, visual perception relies
on topographically mapped representations, but there is no evi-
dence that the representation of objects in working memory is
topographic. Thus, perceptual-level attention might be expected
to have various spatial properties that are absent from working R1. General reaction to commentaries
memory-level attention. A recent study by Vogel (2000) supports
this hypothesis. Specifically, Vogel found that working memory- In BBSs first round of refereeing of the target article, it
level attention can be divided among multiple noncontiguous lo- seemed to several of the referees that the paper was not
cations without being allocated to the regions between these lo- controversial enough to generate useful commentaries. I
cations, whereas perceptual-level visual attention must be focused think that the present set of commentaries has proven to be
on a contiguous region of the visual field. Thus, perceptual-level very interesting, diverse, and thought-provoking after all.
and working memory-level mechanisms of attention differ in both By my count, about 15 of the 39 commentaries solidly
their spatial and temporal characteristics. accepted the hypothesis that some core working memory
It also appears that the mechanisms of attention that operate faculty is limited to about 4 separate chunks of information.
during response selection and initiation are different from those Indeed, at least 8 commentaries (Lane et al.; Nairne &
that operate at earlier stages. The operation of attention during re-
sponse-related processes has been studied by means of the psy-
Neath; Pothos & Juola; Rensink; Rypma & Gabrieli;
chological refractory period (PRP) paradigm. In this paradigm, Todt; Usher et al.; Wilding) presented additional evi-
two stimuli are presented to subjects in rapid succession and the dence for a 4-chunk limit in many circumstances. However,
subjects make separate speeded responses to each stimulus. Re- at least seven other commentaries seemed strongly op-
action time to the second stimulus is slowed when the delay be- posed to that concept and the remaining 17 seemed less
tween the two targets is short, and several studies indicate that this committed either way. Many of the commentators who
is due to capacity limitations at the stage of response selection more or less agreed with the 4-chunk hypothesis offered al-
(Pashler 1994). Pashler (1991) and Johnston et al. (1995) have pro- ternative theoretical explanations.
vided evidence that shifts of visual-spatial attention do not oper-
ate at the same stage as the capacity limits that are observed in the
PRP paradigm. R1.1. Self-justification
In conclusion, there are now many forms of evidence indicat-
ing that there are different mechanisms of attention that operate I believe the target article has made three main contribu-
with different properties within different cognitive subsystems. tions. First, principles were offered as guidelines for iden-
Moreover, it is worthwhile to ask what is gained by proposing that tifying experimental situations in which the number of sep-
the limited capacity of working memory arises from the limited arate chunks can be estimated. Second, a diverse field of
capacity of attention. The real question is why there are limits at evidence was shown to yield similar estimates of capacity in
all, whether we call them limits of attention or limits of working those identified situations; specifically, 3 to 5 chunks on the
memory-specific resources. average, among normal adult humans. Third, various plau-
sible theoretical explanations for the capacity limit were de-
scribed.
The eligibility of evidence often depended upon the ap-
plication of critical theoretical assumptions. For example,
one theoretical assumption provides an answer to Usher et
al., who commented that it is somehow paradoxical that
recall of unattended material should provide a measure for
the capacity of the focus of attention. It is not so paradox-
ical if one considers that, ordinarily, attention exerts effects
broadly across the perceptual encoding, mnemonic storage,
and retrieval processes so that any one particular effect of

Response/Cowan: The magical number 4
attention is difficult to separate from others. By diverting hypothesis that there is a capacity limit of 7 chunks as Miller
attention during encoding and storage phases, the limita- (1956) suggested, and the hypothesis that the number of
tions of attention during retrieval, in particular, can be iso- chunks that can be brought into working memory concur-
lated and identified. Under these circumstances, chunking rently is unlimited. The next section (sect. R3) will be di-
processes are limited and attention can be seen to have a rected at alternative cognitive theoretical accounts of ca-
limit of about 4 chunks of information. pacity limits. A taxonomy of theoretical positions will be
It now seems likely that further, interesting cases of the described; it is this taxonomy that serves as a basic cognitive
capacity limits will emerge. For example, Jinks and Laing metatheory of capacity limits alluded to in the title of
(1999) investigated capacity limits in odor processing, stat- this response. Biological underpinnings of the capacity limit
ing in their literature review (p. 311) that the discrimina- will be discussed (sect. R4) and, finally, I will attempt to
tion and identification of components in odor mixtures is a reach a tentative verdict as to the reason for capacity limits
difficult task for humans, with 34 odorants being the (sect. R5).
limit.
Although I argued for particular fixed capacity limits, I
did not offer a precise, detailed theory of these limits. Some R2. Validity of the 4-chunk hypothesis
commentators (especially Halford et al.; Pascual-Leone;
Milner; Schubert & French) seemed disappointed that The target article has put forward considerable evidence
the rules for determining chunks or storage requirements that, in situations in which a reasonable assumption can be
were not worked out with more theoretical precision, and made about how many separate chunks are in the focus of
that a clear theory of the capacity limits was not proposed. attention, a consistent capacity estimate of 3 to 5 chunks
I agree that such clarity would be desirable if the resulting emerges. Naturally, this evidence hinges on the ability to
theory were correct. However, I disagree with these com- identify chunks correctly. This section first presents a para-
mentators prescriptions for reaching that goal. In the 45 or ble that helps to illustrate the nature of controversies like
so years since Millers (1956) article, this may be the first ex- the 4-chunk hypothesis (sect. R2.1). Next, the key issue of
tensive discussion in which an attempt is made to consider how to identify chunks is taken up (sect. R2.2). This is fol-
the relevance of the burgeoning cognitive literature, taken lowed by a detailed discussion of alternative hypotheses
whole, to the notion of capacity limits. It seems most im- about capacity limits (sect. R2.3). Finally, there is a discus-
portant now to sort out and compile the basic messages sion of a new method for the quantification of capacity lim-
from diverse paradigms in order to provide guidance for its in visual array comparison tasks (e.g., Luck & Vogel
theory, and it would seem unwise to attempt too much the- 1997), which is capable of adding considerable strength to
oretical precision at the same time. the theoretical framework.
Many of the commentaries offered theoretical views that
were more precise than my own but those theoretical views
R2.1. Capacity limits and methods in science: A parable
conflict with one another (as discussed in sect. R4), and
therefore cannot all be correct. If I had foreclosed on a par- Naturally, a capacity limit would be only one of several fac-
ticular theoretical explanation of the 4-chunk hypothesis, I tors influencing memory performance; another, for exam-
suspect that the empirical generalization would tend to be ple, would be the contribution of information that can be
ignored as soon as the particular theoretical explanation of replaced into the focus of attention from long-term mem-
it was disproved, as it most likely would be. Therefore, I ory (i.e., the information within each chunk). Such other
staunchly continue to view the theoretical indecision of the factors will cause apparent departures from the 3- to 5-
target article as a strength, not a weakness. It took 45 years chunk limit that certainly allow room for doubters. It might
for the field to advance from the notion of chunking as an prove helpful to provide an analogy from an area of physics,
aid to remembering, given an unknown capacity limits illustrating how one would expect the 4-chunk hypothesis
pegged at about 7 items (Miller 1956), to an earnest quan- to be criticized even if it were true. Such an illustration can-
tification of those limits and the generation of some testable not prove the hypothesis, of course, but it can at least en-
explanations of the limits. If the result of this article is a new courage further research designed to identify and explain
surge of research and theoretical debate on the reasons for capacity limits.
capacity limits, I believe that the field will be well-served. Thus, consider the case of a now-proven concept in phys-
If a fundamental branch of the tree of possible hypothe- ics, gravity, and how it might have become established in
ses (Platt 1964) could be ruled out, that would be the science (though many of these historical details may be fic-
soundest gain that reasonably could be made at this time. titious). Everyone could plainly see that most objects thrown
Toward this end, a taxonomy of capacity theories will be into the air fall down again. Moreover, some saw that objects
suggested as a metatheory of capacity, and I will consider fall at an accelerating rate that could be roughly estimated
how theories might be compared and assessed. even though the correct equation was unknown. Although
most scientists informally talked of gravity-related ideas,
they could not measure gravity exactly. Some argued that it
R1.2. Organization of the responses
is not a law because some objects do not come down (stars
to specific commentaries
and planets) or come down at much slower, unpredictable
Considering the commentaries can take us a fair distance rates (birds, bats, flying squirrels, leaves, and dust). Propo-
toward understanding capacity limits: Toward that end, I nents of gravity then admitted that other considerations
will first consider evidence for and against the 4-chunk hy- such as distance (in the case of planets) and air resistance
pothesis (sect. R2). In the process, arguments for alterna- have to be taken into consideration. Formulation of a com-
tive formulations will be considered, including the hypoth- plete model taking into account acceleration, size, distance,
esis that there is a capacity limit of only 1 chunk, the and air resistance was a long way off. Some believed in the

theoretical concept that there was a certain rate of acceler- that have strong associations to one another and much
ation at which objects tend to fall, even though it only pro- weaker associations to other chunks concurrently in use
vides an approximation and there are boundary conditions (mathematically, an equivalence class). To apply that defi-
to its use. Given that some of the boundary conditions were nition without direct measurements of chunks, one needs a
unknown, many people remained agnostic and a few denied task analysis of every procedure. It is difficult to see how it
the existence of gravity as a general rule. It may only be would be possible to do better than to analyze each task in
through hindsight that the concept of gravity has come to a logical manner, taking into account what is known about
seem obvious and universal. We have some detailed mod- the task. The fact is that when I used this case-by-case
els of it now, but perhaps we must wait for a future gener- method, the result was a capacity estimate that remained
ation to determine exactly why gravity occurs. fairly consistent at 3 to 5 chunks across a wide variety of sit-
The known history of the discovery of gravity seems con- uations. This constancy in the results itself lends credibility
sistent with this progression of thinking. Tycho Brahe estab- to the method, except in the eyes of those who suspect an
lished a wealth of observations on planetary movement, an unfair use of the method. They must try it themselves.
arena in which there is little wind resistance unlike earthly
applications of gravity. Johannes Kepler used Brahes ob- R2.2.1. A strategy for research to identify chunks. Clearly
servations to establish regularities or laws of planetary mo- what is needed in the future is an attempt to determine
tion, though without really understanding the general prin- chunks more empirically in many relevant procedures. This
ciples behind the laws; and later, Isaac Newton stood on would be a major shift of emphasis for researchers of work-
the shoulders of Brahe and Kepler to establish more gen- ing memory. Measures of reaction time may be of limited
eral laws of gravitational force. Albert Einstein later refor- usefulness because chunking of two elements, A and B, is
mulated the law of gravitation in a more penetrating man- only one possible reason why these elements are recalled
ner that tied it to other forces in the universe, representing together quickly. Supplementary types of relevant evidence
it as curvatures in space; and surely the final word is yet to that two elements are chunked together may include (1) re-
come. call of A and B consistently near each other, and (2) the high
The history of the memory capacity concept arguably probability of recalling B conditional on recall of A, or vice
could be similar to this. On this topic we may have Brahe, versa. Another promising strategy would be to begin with
and this target article reaches toward Kepler, while merely unconnected units and teach the subjects new chunks (with
speculating about Newton and Einstein. Of course, the learning assessment), so that the chunks are known to the
force of this analogy would quickly break down were I to experimenter before they are used in a working memory
choose a shakier concept (e.g., the concept of ether occu- task.
pying empty space, or of a life force). However, it is worth Kleinberg and Kaufman (1971) carried out one proce-
contemplating that apparent exceptions to a causal princi- dure in which the role of chunking in recall was tackled
ple such as capacity do not automatically rule it out if a plau- rather directly. Subjects were presented with clusters of 13
sible case can be made as to why the exceptions occur. The dots, each forming a complex, symmetrical pattern. The vo-
fact that we do not yet have a complete theoretical model cabulary for a particular group of subjects consisted of 2 or
of capacity (at least, not a confirmed one) does not rule it 4 such distinct patterns. Half of the subjects learned to
out, either. group pairs of patterns together to form larger chunks that
It is with this parable in mind that I would address criti- had verbal labels. Each test array consisted of an 8 3 8 ma-
cisms such as the one Avons et al. made, that The capac- trix of such dot patterns. Recall depended on the exposure
ity estimates cited by Cowan are less consistent than he duration, which also was varied, but asymptotically high
claims. We may not have a clear enough understanding of performance levels were obtained with an exposure dura-
the wind resistance type factors in psychological perfor- tion of 6 or 8 sec. For those exposure durations, recall
mance, but we know enough to support a gravity type amounted to about 35 of the designated chunks on day 2
concept (i.e., the capacity concept). So, for example, when for subjects who received 4-pattern vocabularies, or for
Avons et al. state, If four of the most recently presented subjects who received 2-pattern vocabularies and learned
items exist within the attentional focus, then recall of about to chunk them. Recall was at about 8 alleged chunks for
four of the most recent items would be expected. I do not subjects who received a 2-pattern vocabulary but were not
fully agree. The recall can sink below 4 not only because able to chunk them; but one could well imagine that these
some individuals have a lower span than that (e.g., see Fig. subjects learned to chunk the patterns into pairs even with-
3), but also because some attention may be allocated ear- out training in that condition (there being only 4 possible
lier in the stimulus sequence or elsewhere entirely. This pairs, AA, AB, BA, and BB). On days 34, there was a slight
type of variation is bound to depend on factors such as the increase in performance in all groups, which could indicate
motivational level of subjects, the amount of diligence and further chunking.
vigilance that the task requires, the task complexity, and so Illustrating how much work is left to be done, there may
on. be subtle strategic modes that subjects can adopt to alter
the ways in which chunking takes place. Schneider et al.
failed to replicate an instance of grouping observed by Luck
R2.2. Can we identify chunks?
and Vogel (1997), who found that subjects could recall as
Obviously, if chunks cannot really be identified, one cannot many items in arrays of dual-color squares (composed of
quantify the capacity limit in terms of chunks. Several one color peripherally and a different color centrally) as
commentaries (Beaman; Schubert & Frensch) ques- they could single-colored squares. According to the evi-
tioned the target articles ability to determine chunks. Un- dence of Luck and Vogel, each dual-colored square was
doubtedly, this is a key issue. The theoretical definition of a perceived as a chunk. Schneider et al. instead found, using
chunk offered in section 1.3 was a collection of concepts a very similar procedure, that memory for dual-colored

squares was considerably poorer, suggesting that any such array of items (e.g., the color squares of Luck & Vogel
chunking does not take place automatically. The simplest 1997), it is not difficult to guess colors that were presented
kind of explanation is that subtle differences in instructions, so it is the binding of colors to spatial locations that is criti-
subjects, or materials promoted chunking in one case but cal.
not the other. We do not yet know what they are. Given the importance of binding, a chunk might be de-
Schubert & Frensch emphasized the need for an op- scribed in these procedures as a direct association between
erational definition of chunks, saying, what prevents us adjacent items, which together become bound to only one
from arguing that 4 items held in visual short-term memory higher-order serial position or spatial location in the stimu-
(STM) and 4 legs of a chair have a common causality in the lus array, as a group. For example, in a visually presented
limitation of subjects attentional focus? Obviously, in the letter span experiment, upon seeing F-B-I-C-I-A, one
absence of a detailed model (and not just a model, but a cor- might attach the chunk FBI to a first serial position and the
rect detailed model), one must use educated judgment. chunk CIA to a second serial position. In a color-array
(Beaman put the situation in its proper light by calling the memory task, if one were allowed to study the color array
task of determining the number of chunks unenviable. for some time, chunks would consist of multi-color patterns
Certainly, it is the entire field of investigators interested in (most likely composed of three or four colors) assigned to
capacity limits who must share this burden.) Schubert & spatial locations in the array. Perhaps this type of consider-
Frensch suggested an experimental approach in which one ation will help to provide the narrower definition of chunks
would identify a process or mechanism (such as the atten- that Schubert & Frensch craved.
tional focus) that is essential for various tasks showing an ap- Beaman also noted the use of serial-position-specific
parent capacity limit, susceptible to manipulation of a Fac- scoring and suggested that If capacity as measured by se-
tor X. The key finding would be that manipulating Factor X rial recall studies is indeed 4 chunks, then capacity is for 4
would affect all of the tasks similarly. This is indeed a key chunks plus some extra information connecting those
research strategy. However, I would emphasize that one chunks. No, because the items themselves (presumably
must be careful to distinguish between the 4-chunk capac- the chunks in the relevant procedures) do not have to be
ity limit (that only 4 separate chunks can be held in the fo- held in a limited-capacity store. Stated more precisely, they
cus of attention at one time) and the theoretical explanation are automatically activated or accessible and only their
for this limit. The latter topic will be taken up in detail in binding to serial positions within a list (or spatial locations
section R5 but requires a brief discussion now. According within an array) have to be held in the limited-capacity
to the type of theoretical account that I favored in the tar- store. We (J. Lacey, R. Brunner, J.S. Saults, and I) have pre-
get article, the limit is in the capacity of the focus of atten- liminary evidence against a capacity-limited account of per-
tion itself. If this is the case, X could be a manipulation of formance in a modification of the color-array task (Luck &
the amount of supplementary material that must be held in Vogel 1997) in which it is the presence or absence of a color
the focus of attention during the working memory task. anywhere in the array that is tested.
However, according to an alternative theoretical account, In understanding the demands of another type of task,
there is something about the representations of information running memory span, it is important to realize that serial
of various kinds that prevents the focus of attention from positions in a list are not defined on an absolute basis start-
taking in all of the information. For example, perhaps the ing from the beginning of the list and numbered 1, 2, 3, and
most distinct or prominent 4 or so chunks in the field of so on. If they were, then inter-list confusions would occur
stimuli somehow overshadow the others, making them un- between the same numerical positions of lists differing in
available to conscious awareness. In that case, the same re- length. Instead, relative position seems more important.
search strategy applies but the successful X might be some For example, the last item of a j-item list is most likely to be
manipulation that affects the distinctiveness of chunks in confused with the last item of a k-item list, not with the jth
the stimulus field. item of the k-item list (Henson 1999). Until the relative po-
sition of an item is known, it may not be possible to execute
R2.2.2. Further clarification of chunking: Binding and an efficient grouping process. If the length of the list is
chunking in working memory tasks. Commentator Bea- known in advance then the relative position is known; but
man made the important observation that A number of such is not the case in a running memory span task (Pollack
studies which Cowan takes as evidence for his 4-chunk ca- et al. 1959). It may be for that reason that running span re-
pacity limit were serial recall studies in which items were sults in memory of about 4 items per trial, not 7. Items can-
only marked correct if they were recalled in the correct se- not be grouped efficiently as they can in a regular span task
rial order. Nairne & Neath commented similarly, as did because they cannot be assigned to relative serial positions
Avons et al. I completely agree that this is important in un- until the list has ended.
derstanding the capacity limit. In the experiments of serial
recall, a small number of items were used over and over on R2.2.3. Chunking and scene coherence. The principle of
every trial (e.g., spoken digits in the study of memory for scene coherence (sect. 2.6) adds a bit of complexity to the
unattended speech by Cowan et al. 1999). Presumably, all notion of capacity limits in terms of chunks, but it may be a
of the items remained active in memory (or readily accessi- necessary complexity. Its most basic contribution is that it
ble in the test context). It was instead the binding of items shows how the 4-chunk limit could be compatible with the
to particular serial positions within lists that comprised the subjective impression (and behavioral finding) that only
material that must be held in the focus of attention. It would one channel of information can be processed attentively at
be easy to guess that a particular item was a member of the one time (Broadbent 1958). As noted above, the idea is that
current list, so it was the binding of items to their serial po- one can hold in mind about 4 chunks only if they can be in-
sitions that was most indicative of working memory. Simi- tegrated into a coherent scene; only one coherent scene can
larly, in tasks involving the presentation of a simultaneous be held at one time. It is interesting because it is the oppo-

site of a similarity principle. One might propose that two than about 4 previously existing chunks; not, say, 7 or more
chunks will tend not to reside in the focus of attention at the (at least, not without a reiterative use of attention over time
same time if they are too similar and therefore easily con- to build up larger chunks). In the four-chunk view, the sin-
fused. However, another possibility is that two chunks will gle channel held in mind comprises a small collection of
tend not to reside in the focus of attention at the same time chunks that are newly linked together (see sect. 2.2 in the
if they are too dissimilar and therefore difficult to meld into target article). The force of this thesis will be explained be-
a coherent scene. This is a completely open question that low in relation to a number of ostensible counterexamples
has not been empirically studied, to my knowledge. Bea- to the four-chunk hypothesis that commentators brought
man suggests that the work on rapidly recycled sequences up.
of four sounds (e.g., Warren & Obusek 1972) shows this ef- Baars appealed to the case of ambiguous figures, evi-
fect inasmuch as memory for the serial order is much bet- dence of the inability to hold in mind two alternative orga-
ter when the four sounds are similar to one another than nizations of a stimulus at once. Examples include ambigu-
when they are dissimilar. However, this difference can be ous figures and binocular rivalry. However, these examples
attributed to perceptual stream segregation of dissimilar should be disqualified because the alternative organizations
sounds (Bregman & Campbell 1971; Bregman & Dannen- of the stimulus field are logically inconsistent with one an-
bring 1973) and it remains to be seen if there is also a con- other. This inconsistency should be viewed as a form of spe-
ceptual stream segregation that comes into play even in cific interference. Consider, for example, the Necker Cube,
conditions in which the perceptual factors do not operate. which can be perceived in either of two orientations (but
For example, I can think of no work indication whether an not both simultaneously). If the reason for the limit were
array consisting of a color, a tone, a tactile sensation, and a that capacity is limited to one chunk then it should be pos-
printed or spoken word could be held in mind for compar- sible to perceive only one Necker Cube in one orientation
ison with a second, identical or slightly different stimulus or, at maximum, a field of Necker Cubes all in the same ori-
array. entation (presumable forming a chunk). In contrast to this
Schneider et al. referred to Luck and Vogels (1997) suggestion, though, I was able to demonstrate to my own
finding that four features of a visual object can be held in satisfaction that it is perfectly possible to draw two Necker
working memory as easily as one and suggested that there Cubes side by side and to see them in different orientations
may be a limit of four; that, say, six features per object could at the same time. If this is correct, the inability to hold con-
not be held in this way. I personally would not expect such tradictory forms in mind all at once should not be viewed as
a limit, provided that the features are truly integrable. I a capacity limitation.
would not expect it because I would think that the binding Another common objection to the 4-chunk hypothesis,
of features into objects occurs in parallel once the appro- brought up by Baars, is the case of multiple channels of
priate stimulus location has been attended. If there were stimulation. People can attend to only one channel of stim-
such a limit it would suggest that the possible dimensions ulation at once (e.g., not both channels in a dichotic listen-
of difference themselves have to take up space in the lim- ing task). That certainly seems true if the channels are com-
ited-capacity store, in which case the storage limit might plex messages to be understood. However, each channel
have to be revised upward. It seems difficult to think of rel- then contains multiple propositions that must be processed
evant evidence in the literature to date. and integrated. It is therefore important to discuss task de-
mands in a cautious way so as to be sure not to confuse a
channel with a chunk and, as well, not to confuse a percep-
R2.3. Critiques of alternative hypotheses regarding
tual limitation with a working memory limitation.
capacity
Darwin et al. (1972) presented characters in three differ-
R2.3.1. The 1-chunk hypotheses. Commentators Baars ent spoken channels at once and the task required that all
and McElree & Dosher placed their stock in the opinion three be processed. Recall that the whole-report limit in
matching the common subjective impression, that one can this experiment was about 4, suggesting that more than one
hold in mind only one chunk at a time (see also Taatgen). chunk of information could be held in mind at once. It
These investigators do not deny that some faculty of the could be argued that there are other factors to consider
mind is limited to about 4 chunks, but they consider that to such as output interference, which will be discussed below.
be a separate organizing principle for activated information Nevertheless, this experimental result with relatively sim-
outside of awareness. A variety of evidence was adduced in ple channels illustrates that the results of studies with more
favor of this alternative, 1-chunk hypothesis, which de- complex channels are difficult to analyze theoretically for
serves a detailed discussion. capacity limits inasmuch as we do not know how many
These commentators do not mention the concept of chunks are present in each channel.
scene coherence (sects. 2.6 and R2.2.3), which logically Commentators McElree & Dosher brought up the case
could have made them feel better about the 4-chunk hy- of retrieval dynamics. They presented evidence from very
pothesis. Given scene coherence, the subjective impression interesting probed recall tasks in which the speed criterion
of concentrating on only one thing at a time may be real but was varied so that they could plot the increase in accuracy
deceptive. The global broadcasting function of conscious- of recall as a function of the speed of recall. In an unstruc-
ness (Baars 1988) ensures that all items that are present si- tured list, this retrieval dynamic shows the fastest retrieval
multaneously in awareness are linked together and, by dint for the most recent item and no difference between the re-
of that, are experienced as a single chunk. However, this maining list positions. In a list composed of strings of three
linkage is new and can be accomplished only through the consecutive items from a single semantic category, with a
momentary collection of these items all at once in aware- category shift between strings of three, the retrieval dy-
ness. The true capacity limit can be observed in that the namic was fastest for the entire last category (which was as-
new chunk that is formed can be constructed out of no more sumed to be represented as a single chunk in memory). A

category cue was capable of conferring this advantage on a from a category is either 1 or 3. According to the single-
non-final category. On the basis of these findings they pro- chunk hypothesis, a single ungrouped item in list-final po-
posed that one chunk is in the focus of attention in these sition should be retrieved more quickly than any of the
tasks, resulting in the fast retrieval dynamic. items within a three-item list-final category because the iso-
There are at least three problems with the interpretation lated item avoids an unpacking stage. Failure to find this
of this evidence. First, assuming that the most recently pre- would seem to indicate that items are held in memory as
sented or cued chunk is in awareness, the advantage seen separate (though associated) chunks, not as one chunk.
for this chunk may not indicate that it is the only chunk in Last, there is an alternative interpretation (mentioned by
awareness. For example, in a list of six unrelated items (pre- McElree 1998) in which the advantaged items are not nec-
sumable separate chunks), perhaps the last item is consis- essarily in the focus of attention but are simply more in
tently in awareness and the rest of the available capacity is keeping with the contextual cues of retrieval. If this is the
used up with a rehearsal set that includes a subset of items case, or if the true capacity limit is greater than one chunk,
from anywhere else in the list. Sometimes it would include it should be possible to achieve a faster retrieval dynamic
early items, sometimes it would include medial items, and for more than one category at once. If such a result were
sometimes it would include penultimate items. If this were apparently obtained, though, one would have to look at the
the case, one could well obtain exactly the pattern that was distribution of responses to ensure that subjects did not
observed: a retrieval speed advantage for the final serial po- simply choose one category to keep in the focus of attention
sition compared to all other serial positions, which need not on each trial; this presents a thorny problem to be ad-
differ from one another. dressed in future research.
Second, it may be that subjects do not always fill their ca- McElree & Dosher also mentioned the simultaneous
pacity, sometimes choosing instead to include only one versus sequential array distinction. They suggested that the
chunk in the focus of attention even though they are capa- four-chunk hypothesis may be valid only when the chunks
ble of including more (though perhaps capacity can be are presented simultaneously (e.g., Luck & Vogel 1997).
filled only at a cost to the strategic processing that can be Apparently, they dismissed the sequential evidence, such as
done). the memory-for-unattended-speech procedure of Cowan
Third, and finally, there is a potential problem in the des- et al. (1999). It is not clear why they would consider it in-
ignation of a chunk. Above, chunks were described as sets valid but one could argue that there is output interference
of items with strong intra-chunk associations and weak in- from each item being recalled in that procedure, reducing
ter-chunk associations. Presenting three items from the the ability to recall subsequent items. Indeed, that is a com-
same category does not automatically make them into a plaint lodged by Milner and by Nairne & Neath. An un-
chunk, although it does produce some degree of associa- published experiment (by N. Cowan, J.S. Saults, E.M. El-
tions between them. For example, if the related items oc- liott, and L.D. Nugent) allays that particular concern. Lists
cupy serial positions j, j 1 1, and j 1 2, it is possible that of nine digits were presented in an unattended channel and
within the retrieval context, item j 2 1 (not in the categor- only occasionally were tested, in order to minimize the role
ically related set) is more closely related to item j than item of attention. In this particular experiment, recall was to be-
j 1 2 is, as a result of the adjacency of j 2 1 and j in the list. gin at any of the nine serial positions and was to continue to
If one rejects the notion that the categorically related items the end of the list and then starting back at the beginning,
form a chunk, the alternative account is that the associa- in a circular manner, until all nine items were recalled. One
tions between items are enough to induce subjects to hold can look at the proportion correct for any of the nine serial
all three of the items from a category in mind indepen- positions for the first-recalled items, for which there was no
dently. One critical prediction of this alternative account output interference. The results examined in this way were
would be that one could not get the enhanced retrieval dy- bow-shaped with a relatively weak primacy effect and a
namic with, say, seven items in a row from one category fol- stronger recency effect. Summing the proportion correct
lowed by seven from another category because the items for first-recalled items across serial positions produced an
from a category then could not be kept in the focus of at- estimate of the number correct in the absence of output in-
tention concurrently. terference. This sum came to 3.2 items, consistent with the
It is not altogether clear how performance should be af- 4-chunk hypothesis as elaborated above. There at least does
fected if the three items from the most recent category not appear to be any strong evidence against the hypothe-
were held in that focus of attention as a single chunk. One sis that attention focused on auditory sensory memory can
would think that this would provide strong enough infor- result in the apprehension of roughly 4 chunks of informa-
mation to distinguish members of a category that were pre- tion. Thus, this study suggests that one can extract 4 chunks
sented, from other similar members that were not pre- from a sequential presentation. In principle, moreover, it
sented and therefore were not part of the chunk. Yet, for should be possible to use ones faculties of mental imagery
the most recently-presented category, false alarms to non- to recast sequential arrays as simultaneous arrays.
presented members of the most recent category (e.g., re- In partial defense of the suggestion by McElree &
ceiving pig, dog, and mouse and false-alarming to cat) oc- Dosher that the 4-chunk limit could apply only to simulta-
cur relatively frequently (McElree 1998, Fig. 4). Another neous presentations, though, one possibility is that attended
possible prediction is that, if the three recent members of sequential presentations do not typically result in the load-
a category really do form a single chunk in the focus of at- ing of the focus of attention with four consecutive chunks
tention, there should not only be an advantage arising from of information. Instead, perhaps rehearsal processes (of an
their presence within that focus; there also should be a dis- elaborative nature?) use up some of the capacity. One thing
advantage arising from the need to unpack the chunk in or- that favors this possibility is that, as pointed out by Nairne
der to access its individual members. This could be tested & Neath, there is a study that contradicts the finding of
in a procedure in which the number of consecutive items Halford et al. (1988), discussed above, indicating that

proactive interference does not occur for a memory test in lags varying from 1 item to 16 items. For this sort of list, it
which the target list is only four words long. Tehan and is impossible to assign items to specific serial positions, so
Humphreys (1996) did find such proactive interference. grouping and rehearsal become difficult or impossible. The
There are numerous differences between the studies that results show facilitated (fairly accurate) lag judgments for
need to be explored but one potentially critical difference lags of 1 through 4, with an asymptotic average judgment of
is that Halford et al. presented the words within each list as about 57 items applied indiscriminately to items with an
a simultaneous array, whereas Tehan and Humphreys pre- actual lag of 516. It is as if the capacity-limited processing
sented each list sequentially. It may be that, because of this system is capable of noting only 1, 2, 3, 4, and many.
difference, 4 chunks entered the focus of attention at once It is going to take additional work before we will under-
in the Halford et al. study but not in the Tehan and Hum- stand exactly what testing procedures best index informa-
phreys study. (However, the situation may be different in tion in the focus of attention. Oberauer (in press) has de-
free recall, where proactive interference is not found for vised a modified probe reaction time testing method that
the recency portion of the serial position curve; see Craik & seems very promising for distinguishing between informa-
Birtwhistle 1971.) tion that is in the focus of attention versus information that
Another comparison of hypotheses has to do with the ca- is activated, yet outside of the attentional focus (Cowan
pacity and function of the focus of attention and with how 1988; 1995; 1999). Oberauer presented two lists of items
comparisons of chunks are carried out. For example, in a (digits or words in different experiments) concurrently on
probed recognition task like the one that McElree & each trial, in different colors. The set sizes of both lists var-
Dosher often use, it is necessary to compare the various ied. A postlist cue indicated which list would be tested. Af-
targets to the probe until a match is found or none can be ter a variable interval, a recognition test probe followed the
found. If only one chunk can be held in the focus of atten- cue. The measure of the irrelevant list remaining activated
tion at a time, that implies that the probe and target cannot in memory was an intrusion effect: specifically, the slowing
both be held in the focus of attention at the same time. effect that occurred when a negative probe was an item
Therefore, the comparison process must take place outside from the irrelevant list rather than a nonpresented item.
of that focus. It is difficult to understand why a process However, the measure of the irrelevant list remaining in the
should be strategically controlled and yet take place outside focus of attention was modification of the intrusion effect
of attention, and why the probe itself should not be consid- by the set size of the irrelevant set. The intrusion effect
ered part of the focus of attention at the same time as at lasted a long time and was larger for older adults, whereas
least one target item. Similarly, the global workspace idea the irrelevant set size effect occurred only for several hun-
mentioned by Baars seems useful largely because it pro- dred milliseconds, presumably until the irrelevant list could
vides a mechanism whereby independent chunks can come be removed from the focus of attention, and did not differ
together within the focus of attention to form a new chunk with age. In another experiment, in which subjects (young
that reflects the present episode (a theory that also provides adults) knew that both lists had to be recalled, the results
a teleological argument for the between-object links hy- reassuringly suggested that both lists remained in the focus
pothesized by Davis and consistent with the episodic of attention. This procedure could be used to estimate the
buffer of Baddeley). If multiple chunks cannot be present contents of the focus of attention in a more principled man-
in the focus of attention long enough for them to be linked ner than has been done previously.
together, it is not clear how or where in processing this
chunking process can be carried out or why newly formed R2.3.2. The 4-chunk perceptual/1-chunk memory hy-
chunks are limited to a combination of about 4 pre-existing pothesis. Commentator Jou posed in interesting criti-
chunks (see sect. 2.2 in the target article). cism. If the memory search procedure of Sternberg (1966)
McKone reviewed results that appear to be at odds with and the enumeration procedure (e.g., Mandler & Shebo
those of McElree and Dosher. Recognition judgments were 1982) both reflect the focus of attention or window of si-
to be made for items presented at various points within a multaneous consciousness (Ebbinghaus, cited in Slamecka
long list; the reaction times were faster for the most recent 1985), then why do they show such different patterns of re-
4 items, in keeping with a prediction made in the target ar- sults? Specifically, memory search tasks yield a linear in-
ticle. However, this experiment also showed that the reac- crease in reaction times as a function of the set size up
tion times were shorter for stimuli presented more recently through about 6 items and then a shallow slope, whereas
within the last 4 lags. This suggests that the more recent enumeration tasks yield a shallow slope for 1 4 items and
items were fresher in mind, not equally accessible within a a steep slope after that. Jous suggested explanation ap-
focus of attention. The implications of this finding will be pears to be that, in perception, the simultaneous presence
discussed further in section R2.3.5. However, for now it is of several stimuli results in parallel processing of the items
worth noting that more work must be done to reconcile in the focus of attention; whereas, in memory search, items
these findings. McKones experiment did not examine re- enter the focus of attention only one at a time in a serial
trieval dynamics (changing accuracy as a function of research process.
sponse time criterion), so it cannot be directly compared to There are several points that must be made in response.
the test situation described by McElree and Dosher; but First, it seems unlikely that memory search occurs in a
the two experimental procedures may well differ in how straightforward, serial fashion. It has often been found that
much rehearsal they allow. In McElree and Doshers situa- search times are faster for items at the end of the list than
tion, rehearsal may displace some specific items (e.g., the for items earlier in the list (Clifton & Cruse 1977; Diener
three penultimate items) from the focus of attention. 1990; Dosher 1981; Monsell 1978; Ratcliff 1978), which ap-
Data basically similar to McKones can be seen in Atkin- pears to rule out a serial search of items in the presented
son and Shiffrin (1968, Fig. 24). Lag judgments were to be serial order. Although it is possible that memory search oc-
made for stimuli presented in a long, running list, with test curs in some random or backward order, it seems more

plausible, and there is better evidence, that memory search R2.3.3. The 7-chunk hypothesis. Two commentators
occurs for list items in parallel, in a capacity-limited man- (Bachelder and Pascual-Leone) found value in the 7-
ner (Ashby et al. 1993; Ratcliff 1978; Townsend 1976), with chunk hypothesis, although presumably for very different
recent items often having an advantage within the compet- reasons.
itive process because of their greater salience in memory. Bachelder showed that points from several types of
Second, the difference between reaction time patterns is experiments (memory span, subitizing or span of appre-
not the result of a contrast between perceptual tasks on one hension, and absolute judgment) could be plotted on a sim-
hand and memory tasks on the other. In visual search tasks ilar, inverse-ogive function (as in his figure). He suggested
with a varied target set from trial to trial, which are per- that the magical number 4 emerges from a perfect-
ceptual tasks, one sees a reaction time that increases stead- performance criterion on this function, whereas Millers
ily with set size, not the two-part pattern seen in subitizing magical number 7 emerges from a criterion at which 50%
(Dosher 1998; Schneider & Shiffrin 1977). of the trials result in correct responses. (At least, I assume
Third, one must take into consideration differences in that is what the curve shown in Bachelders figure repre-
task demands. In enumeration, the entire pattern of objects sents.) Although this function is empirically beautiful, I
is relevant to the count. In memory search, in contrast, what have some doubts about its interpretation, for several rea-
is needed is not a conglomeration but a discrimination sons. First, correct performance on a memory span trial re-
among items. In one way or another, the probe must be quires success at all serial positions. In contrast, the span of
compared to each item in memory at least until a match is apprehension and absolute judgment tasks required just
found. For any memory set size greater than one, the total one response per trial.
collection in the focus of attention will not match the probe. More important, there is no way that the curve shown in
There is no reason to think that the process of determining Bachelders figure can represent all of the evidence that I
whether one of the items within the focus of attention take to reflect capacity limits. Consider, for example, the 35
matches the probe should be automatic, so it will depend adults who carried out the memory-for-unattended-speech
on the set size. task of Cowan et al. (1999). These subjects completed at-
Fourth, it is easily understandable from a fixed-capacity tended-speech as well as unattended-speech tasks. The re-
view that the supra-capacity slope differs markedly from sults corresponding to Bachelders plot (proportion of trials
one task to another. Specifically, whether the reaction time correct at each list length) are shown in Figure R1. The bold
function at large set sizes becomes steeper (as in enumera- solid line is a portion of Bachelders normal ogive. The plain
tion tasks) or flattens (as in memory search) clearly must de- solid lines with solid points represent data from the at-
pend on what process takes over at that point. In the case tended-speech procedure. Each subject received trials at
of enumeration, it is a serial counting process, whereas in only 4 list lengths, equal to the longest list that was recalled
the case of memory search, it is a parallel direct access correctly in a pretest (defined here as span) and three
process that is not capacity limited, yet is slower than the shorter list lengths (span-1, span-2, and span-3). To obtain
capacity-limited process that one finds at smaller set sizes. a wider range of list lengths, data are shown separately for
So this difference between findings is not a problem for the subjects with span 5 6 (N 5 5), span 5 7 (N 5 14), span
4-chunk hypothesis. 5 8 (N 5 11), and span 5 9 (N 5 5). The ogive approxi-
Despite these points, it must be acknowledged that a mates these functions fairly well; it seems nearly perfect for
central riddle remains unanswered. Why should the reac- subjects with a span of 6. However, look now at the dashed
tion time in memory search increase linearly as the set size lines, representing data from the unattended-speech pro-
increases to 6 items if only 4 can be held in the focus of at- cedure. In this procedure, scores are considerably lower. In
tention? That is a good question. Nevertheless, it is worth fact, for these data, it is approximately the case that the 50%
pointing out that subjects in memory search tasks are al-
lowed to rehearse the stimuli. The process of rehearsal may
keep up to 6 or so items in such an active state that the re-
sponse times remain faster than is found through capacity-
unlimited memory access. According to this suggestion, it
would be predicted that the inability to rehearse during
presentation of the memory list should shift the point of dis-
continuity leftward, so that the linear increasing portion of
the function should extend through set sizes of 4 but no
more.
This still does not really address the question of why re-
action times increase in a linear fashion through 6 items in
memory search. That point has been a problem, though, for
all theories of memory search. We know that various types
of serial or parallel capacity-limited processes can produce
linear functions but we do not know why the functions hap-
pen to be linear despite nonlinear serial position effects in Figure R1. Proportion of lists correctly recalled by the 35 adults
who carried out the task of Cowan et al. (1999); these are the same
reaction time. I am confident that it would be possible to subjects shown, for example, in Figure 3 of the target article. Here
put forth a theory of the linear function that assumes a core they are reported separately for subjects with each memory span
attentional capacity of 4 chunks but, given how little is un- (defined as the longest list repeated correctly in a pretest). Thin
derstood here, it would be post hoc and unhelpful to do so. solid lines, attended-speech condition; dashed lines, ignored-
Instead, I would point to this as an exciting area for further speech condition. Only the attended-speech condition matches
empirical research in the near future. the normal ogive (bold solid line). Based on data by Bachelder.

correct criterion occurs for 4 items, not 7. This may reflect logical constraints that apply similarly to both STM and
the fact that the estimated capacity is 4 chunks or higher LTM. For example, the explosion in the number of ways in
only in about half of the adults, as shown in Figure 3. which one can arrange n objects as n exceeds 4 could add a
A devils advocate could point out that one measure taken difficulty to both STM and LTM mechanisms even if they
to indicate a pure STM capacity limit, the subitizing (span operate separately.)
of apprehension) limit of Mandler and Shebo (1982), fits Pascual-Leone was more explicit in arguing for a 7-
Bachelders function, whereas another measure, the unat- chunk hypothesis. The basis of the argument was that the
tended speech procedure of Cowan et al. (1999), does not, analysis in the target article was incomplete and did not take
suggesting that they cannot represent the same concept. I into account the memory capacity needed for the proce-
would argue, though, that this discrepancy can be explained dures involved in carrying out the task. However, this analy-
within the fixed-capacity view through a theoretically- sis would be invalid if, in fact, the procedures have become
oriented task analysis. Memory span and absolute judg- sufficiently automatized so that they do not compete with
ment, which I consider to be compound STM estimates, limited-capacity storage. A case in point is Pascual-Leones
could both require rehearsal of a stimulus series (the list analysis of the unattended-speech task of Cowan et al.
items for span and the response choices for absolute judg- (1999). According to that analysis, 3 or 4 slots in working
ment). In subitizing, however, all items are considered at memory are used up to carry out the rhyme-matching
the same time and only one response is to be emitted. Loss primary task, and a remaining 3 or 4 are used to retain in
of some precision of information may or may not throw off memory digits from the unattended channel, which are rec-
the single response, whereas loss of precision in a span sit- ognized following an orienting response. This analysis con-
uation is likely to result in one or more incorrect item re- trasts with the analysis of Cowan et al., who assumed that
sponse, making the entire list response count as incorrect the digits were held automatically in a sensory memory and
for the type of data shown in Bachelders figure. Thus, the that attention (and, therefore, capacity-limited storage) was
empirical functions shown by Bachelder do not consistently focused on the speech only after the rhyming task was re-
indicate the pure capacity limit. placed by the auditory-memory recall cue. One problem
There is, nevertheless, a problem in understanding subi- with Pascual-Leones analysis is that Cowan et al. found no
tizing according to the 4-chunk hypothesis. If only 4 items effect of the concurrent memory task on performance in
are apprehended as separate chunks and the other items in the rhyme-matching task. The same was true of Cowan et
an array are not apprehended, shouldnt performance be al. (2000), who used the same task but manipulated the
approximately a step function of items, with the step oc- memory retention interval rather than the list length. Also,
curring at the capacity limit? At least two alternative expla- if working memory slots were used up retaining some of the
nations of subitizing have been introduced. As mentioned supposedly unattended sounds, one might expect the ca-
by Taatgen, Peterson and Simon (2000) suggested that pacity to remain fairly constant across a silent retention in-
rapid enumeration of small numbers of objects comes from terval. This would not be expected if performance depended
the recognition of specific patterns that the objects happen on a sensory memory that was lost during the retention in-
to form, and showed that the number of patterns to be terval. A reanalysis of adult data from Cowan et al. (2000)
learned begins to skyrocket for numbers above 4. Avons et shows that the capacity (and its SD) was 3.34 (1.53) after a
al. Suggested that numbers of objects are judged in a con- 1-sec retention interval, 2.27 (0.85) after a 5-sec retention
tinuous manner and that the just-noticeable difference interval, and 1.90 (0.73) after a 10-sec retention interval.
(jnd) between two amounts depends on the Weber frac- This is in contrast to performance in the attended-speech
tion. (Milner made a similar suggestion.) The idea would condition, which showed no memory loss across the reten-
be that for a 4-item array, the Weber fraction is small tion intervals. Thus, the evidence is more consistent with
enough that the true number plus or minus the jnd is still the interpretation of Cowan et al. (1999), involving a 3- to
nearer to 4 than to the neighboring numbers; whereas for 4-chunk limit drawing upon a sensory memory representa-
arrays of 5 items or more, the jnd is large enough for the tion of the last spoken list at the time of the recall cue.
number to be mistaken for a neighboring number. This is It also would be difficult for Pascual-Leones model to
an interesting possibility, although it is unclear if the mag- interpret one finding of Luck and Vogel (1997) in their
nitude of differences could account for the discontinuity color-array memory task: that performance was unaffected
observed between 4 and 5 items. For example, according to by a concurrent, 2-item, verbal memory load.
an application of Webers law, suppose that the just-notice- How, then, would we account for the evidence upon
able difference for 4 objects was .44, so that the perceived which Pascual-Leones M-space theory is based? One pos-
number could not reach any lower than 3.56 or any higher sibility to consider is that the tasks he used allowed a con-
than 4.44. In that case, the predicted jnd for 5 objects would tribution of rehearsal and chunking strategies to improve
have to be (5/4) 3 (.44), or .55. With some random error performance, providing compound STM estimates rather
thrown in, it is unclear whether such a subtle change in the than pure STM capacity estimates.
jnd could truly account for the discontinuous performance Kareev (2000) offered another argument in favor of a
function. One prediction of the model, however, is that the magical number 7 6 2. The argument was based on con-
point of discontinuity should increase markedly if the task siderations about correlations between binary variables.
requirement is relaxed (e.g., not asking for an exact count Small samples tend to overestimate the correlation found
but asking whether there were ,X objects or not). At pre- in the entire population. Given that a smaller working
sent, I am far from certain that the limit in subitizing occurs memory provides an individual with smaller samples, a limit
for the same reason as the other capacity limits reviewed in on working memory could be useful for detecting weak cor-
the target article. This does not, however, provide any sup- relations (see sect. 4.1.1). Kareev (2000) argued that the
port for the 7-chunk hypothesis or weaken the 4-chunk hy- usefulness index was greatest for samples of around 7 pairs
pothesis much, generally. (It also is possible that there are of values. Although this might be viewed as embarrassing,

given that the target article used the same phenomenon to Ericsson & Kirk. They stated that Even in task environ-
argue for a capacity limit of 4 rather than 7, I stand by my ments where the functional independence of chunks is con-
estimate, on the basis of all of the evidence taken together. vincingly demonstrated, individuals can increase the stor-
The arguments based on correlations are lax enough to age of independent chunks with deliberate practice well
allow either estimate. An argument demonstrating the above the magical number four. One may well question,
plausibility of 4 emerges from a rudimentary simulation I however, what the effects of expertise are and what was
have carried out with slightly different assumptions than meant by the functional independence of chunks. For ex-
Kareev used. I created a random list of 1,000 numbers (x1, ample, in the memory performance of a trained individual
x2, . . . x1000) with the value of 1 or 2. I then created an- who can repeat lists of many digits in a row (Ericsson et al.
other list of 1,000 numbers (y1, y2, . . . y1000) with the 1980), there may be a hierarchical organization in which
value of 1 or 2 so that the correlation between the two lists chunks are composed of other chunks. It is the higher-level
equaled .4. A third list was constructed, so that its correla- chunk that should govern recall ability (e.g., features are
tion with the first list was .6. For each pair of correlated lists chunked into letters but capacity is not measured at the let-
(12 and 13), a moving window of 2, 4, 6, or 8 number ter level when letters are in turn grouped into words).
pairs was used to calculate sample correlations. If there was Ericsson & Kirk noted two encoding mechanisms by
no variation in the sample obtained from at least one of the which information can be stored beyond any specific ca-
lists, no correlation was calculated. A particular value of the pacity limit. They were said to be able to generate associ-
sample correlation was taken as the minimum criterion (C) ations between different presented chunks of information
indicating that the subject would notice the correlation, and build new integrated structures in LTM. It was noted
and the results were examined with various criteria. The that If experts can encode associative relations between
question then was, which number-pair sample set size (2, 4, virtually all chunks within their domain of expertise, then
6, or 8) produced the largest number of noticed correla- the concept of chunk independence would not apply. Sec-
tions in the same direction as the population correlation? ond, and more important, skilled individuals acquire skills
The number of noticeable correlations was corrected by to associate chunks of presented information with internal
subtracting from the number of correlations with r $ C the cues organized in retrieval structures. I do not deny any of
number that were noticeable, but in the wrong direction (r this. However, if experts are able to do this they may be able
# 2C), to produce an adjusted success value. The outcome to represent an entire stimulus field as a single chunk that
was similar for a population correlation of either .4 or .6. can be unpacked (with the help of long-term memory) into
With C set to .9, set size 4 was clearly the winner. (With a smaller chunks as needed. Therefore, the ability of experts
population correlation of .4, the set sizes of 2, 4, 6, and 8 to overcome the conventional capacity limits can be viewed
produced adjusted success values of 198, 212, 114, and 57, as an extraordinary instance of the capacity limit but not an
respectively. With a population correlation of .6, the same exception.
set size produced adjusted success values of 290, 352, 263, An important complication for future research is that
and 182, respectively.) Set size 4 also consistently came out some associations may be intermediate in strength. So, for
ahead with C set to .6, whereas set size 6 won with C set to example, if one were asked to recite the alphabet and then
.7 or .8, and set size 8 won with a C set to .5. Given that we the American pledge of allegiance to the flag, one would
do not know the appropriate detectable sample correlation pause briefly within the alphabet because some groups of
level C for human subjects, this exercise yielded no clear-cut letters form a subgroup; and one might pause during the
winner. It suggests that in our present ignorance, one should pledge because a certain phrase is not immediately acces-
not take Kareev (2000) as strong teleological evidence either sible. Nevertheless, if we can learn to understand effects of
for or against a basic capacity limit of 4 chunks. intermediate-level associations and of grouping hierarchies
Wilding presented protocols of mnemonists who could used in multiple reiterations, it is reasonable to apply a ca-
recall long strings of memorized digits. Reaction times pacity model. Basically, all one has to do in the aforemen-
showed breaks after sets of about 4 digits but, in an appar- tioned situation is to hold in mind a node for the alphabet
ent departure from the 4-chunk hypothesis, the last series and a node for the pledge of allegiance, while unpacking
could be as long as about 7 digits. One might suspect that a and recalling subgroups of elements one by one. The more
decay-based mechanism held the last string of 7 or so dig- expertise is involved, the more theoretical work investiga-
its until they were available for entry into the focus of at- tors must do to understand the associative structure before
tention. However, according to the theoretical view sug- a capacity principle can be applied. This, however, does not
gested in the target article, it is mysterious how the focus rule out the principle. Ruling out the existence of capacity
of attention could collect and recall the last 7 in a single limits on the basis of the phenomenon of expertise would
cluster. Close inspection of Wildings Figure 1 suggests that be much like ruling out the existence of controlled pro-
there actually may have been slight pauses in the middle of cessing (Shiffrin & Schneider 1977) on the basis of the phe-
the last group of 7. They could be difficult to detect if the nomenon of automatic processing, or more broadly, like
retrieval process for the last chunk could take place largely ruling out effects of gravity on the basis of effects of wind
during the recall of the next-to-last chunk. This could occur resistance.
at the end of the list because the overall structure of the list Schubert & Frensch made an interesting suggestion
no longer had to be held in mind. Another possibility is that along this line. They suggested that in dual-task studies a
the last two chunks of 3 or 4 numbers were melded into a capacity limitation of the central mechanism can be ob-
single super-chunk in the period after the list was presented served only when subjects carry out the tasks in a relatively
and before recall. unlearned state. They go on to suggest that training leads
to an over-learned mapping of stimuli and responses, and,
R2.3.4. The unlimited-capacity hypothesis. A completely consequently, to an automatic activation of the response
different view from the present target article was taken by when a stimulus is presented. In other words, expertise

uses a type of automatic processing that can obscure the ef- recalled (e.g., the single-decision procedure of Luck and
fects of capacity limitation within controlled processing. Vogel 1997; the unpublished paper on unattended speech
The task for a capacity account would be to determine what discussed in sect. R2.3.1, in which just the first-recalled
little controlled processing is left to be accomplished by the item was examined). It also is worth recalling evidence that
expert in a capacity-limited manner. For example, a chess information in the focus of attention may be impervious to
expert may be able to go around a room playing multiple some types of interference that can occur outside of that
chess games at once, but it seems doubtful that chess ex- focus (Craik & Birtwhistle 1971; Halford et al. 1988, dis-
pertise would allow an individual to play one game with the cussed in sect. 3.3.4). Therefore, I believe that the concepts
left hand and a different game with the right hand, without of time-based forgetting and material-specific interference
some loss of speed or playing quality. I, thus, adhere to the are more appropriate for describing passive storage buffers
principle that expertise is likely to alter the way in which a (Baddeley 1986) or activation outside of the focus of atten-
limited capacity can be filled, but without doubting the cation (Cowan 1988; 1995). To understand short-term mem-
pacity itself in terms of chunks (or a more sophisticated, ory, one must know about the capacity of attention (in
graded, hierarchical associative structure once it is under- chunks) as well as the passively held information available
stood and studied). to it (affected by time and interference). The intent of the
One commentary (Lane et al.) reviewed work on com- target article is to improve our understanding of short-term
plex tasks such as chess, in which, following Chase and Si- memory by analyzing a key component of it separately but
mon (1973), they have been able to obtain empirical evi- we will then have to put the whole working memory system
dence of chunks. They were able to implement a specific back together. (For attempts to do so see Cowan 1999; also
computational model of the process and found that a 3- or see the commentary by Baddeley and see Baddeley, in
4-chunk capacity best matched performance of players at press.) Additional discussion will be in response to the spe-
different levels of expertise. Although this theoretical suc- cific points of each commentator, who used the notions of
cess is in only one research domain, it is a reassuring decay ind interference in different ways.
demonstration that such work can be done. Davis described an unpublished experiment in visual
perception by Davis et al. that he thought ruled out a
R2.3.5. Decay-and-interference-only hypothesis. Several chunk-based capacity limit in favor of decay and interfer-
commentators (Davis; Grondin; McKone; Milner; Mu- ence. There was not much actual emphasis on decay. The
ter; Taatgen; Towse) suggested that the possible influ- interference that he mentioned apparently referred to a
ences and results of limiting factors other than chunk limit in how many binding links could remain activated, be-
capacity (in particular, time-based forgetting and interfer- tween objects and between features within an object. That
ence) were neglected or dismissed. For example, Muter type of interference, if correct, could serve as a subatomic
suggested that a comprehensive account of STM should type of analysis of capacity limits that is essentially friendly
surely include treatment of the nature of forgetting after at- to the notion of a 4-chunk limit, which he predicted on the
tention has been diverted. I could not agree more and ad- basis of the number of links for ordinary objects, in which
dressed those questions at length in several previous works the total number of inter-object links increases rapidly with
(Cowan 1988; 1995). I have studied time-based forgetting the number of objects in the display.
intensively in the past (see for example Cowan 1984; 1988; Although this reasoning is exciting, the conclusion that
Cowan et al. 1990; 1994; 1997; 2000; in press; Gomes et al. Davis described on the basis of an experiment seems open
1999; Keller & Cowan 1994; Keller et al. 1995; Saults & to question. In this experiment, subjects had to compare the
Cowan 1996). Time-based forgetting has been difficult to shapes of two notches among 3 or 6 irregularly-shaped ob-
understand inasmuch as, after all these years, there still may jects, as shown in Parts C and D of his figure. The number
be no clear evidence for forgetting as a function of the pass- of objects (3 or 6) did not matter overall, but one case was
ing of a particular, absolute amount of time as opposed to an exception. In that case, horizontally displaced notches
the passage of time relative to the times of presentation of appeared on the same object in either situation, whereas
potentially interfering stimuli (e.g., see Cowan et al. 1997 vertically displaced notches appeared on the same object
vs. Cowan et al., in press). Nevertheless, time-based for- only in the case of 3 objects, not in the case of 6 objects. In
getting and interference from other items are both very im- this case, there was an advantage of fewer objects. This ex-
portant factors to consider in recall. The present target ar- ception showed that the number of perceived objects in the
ticle is simply intended to cover only one theoretical displays really did differ as intended. Thus, the absence of
mechanism of short-term memory, the capacity limit. a number-of-objects effect overall was seen as evidence
Towse noted that Strong faith in measures of memory against a capacity limit of 4 objects. The ordinary effect of
size alone may permit us to find The Answer, but at what number of objects was said to be an indirect effect of the
cost to a full understanding of immediate memory? The number of inter- and intra-object links combined, which
implication is that he believes that chunk-based capacity here were held constant. My concern with this logic is that
limits may eventually turn out to result indirectly from de- the demands of the task are not clear. Not all visual searches
cay and interference factors. Although I believe that decay are capacity-limited; simple feature searches are indepen-
and interference are critical for remembering, their useful- dent of the number of objects. If the two target notches in
ness seems reduced when one is talking about the contri- the display can be located in a capacity-unlimited search,
bution of the focus of attention, the topic of the target arti- then there is no reason to expect that the comparisons
cle. What is crucial here is that the 4-chunk capacity seems would be made more slowly with 6 objects present than with
to apply even in situations in which conditions have been 3. Also, the more specific effect that was obtained could
selected so as to minimize time-based forgetting of infor- have occurred not because of a difference in the number of
mation before it reaches the focus of attention and output objects perceived in the two types of display but because of
interference that could conceivably displace it before it is subjects use of local symmetry as a cue. It may be especially

easy to compare two notches when they flank a large, sim- during a secondary task, for example, remembering to
ple object because, in that circumstance, the notches are check the oven in a timely fashion while busy composing a
the same if and only if there is local symmetry in the design. journal article. Despite the importance of not burning
For the horizontally displaced notches, symmetry is always down the house, the stove presumably cannot remain in
a useful cue because a rectangular edge is flanked by the working memory and must be reloaded periodically, most
notched (though perhaps this cue is not as useful in the case likely during breaks in the mental workload of the ongoing
of 3 large objects because the horizontal symmetry of the composition.
semi-elliptical shapes could make it difficult to perceive lo- McKone presented results that were discussed briefly in
cal asymmetry). For vertically displaced notches, in the case section R2.3.1. and are relevant to the decay issue. Recog-
of 3 large objects, the notches flank a white semi-elliptical nition times were found to be shortest to the most-recently-
shape; that design feature might be used to judge symme- presented item and to increase steadily until a lag of 4, af-
try. For 6 small objects, in contrast, vertical displacement ter which the response times stay much more constant.
leaves no large shape for the notches to flank, which may Given this pattern, she noted that it conflicts with the idea
make symmetry less useful. Daviss point is potentially very that the most recent 4 items are equally well represented in
interesting but it seems that these possibly confounding fac- the focus of attention. There, she suggested that older
tors first need to be examined further. items fade from the limited capacity mechanism (at least
Grondin brought up an interesting temporal limit in cog- without rehearsal). What makes me doubt that interpreta-
nition. The comparative judgment of time intervals benefits tion is the pattern of results from our experiments on mem-
from rehearsal only, according to his findings, for intervals ory for unattended lists of spoken digits (e.g., Cowan et al.
longer than about 1.2 sec. Presumably, this is an estimate of 1999; 2000b). These experiments show a distinct bow-
the psychological present, which could be described as shaped serial position function. One interpretation is that,
the period for which stimulus input seems to cohere into the at the time of the recall cue, the focus of attention casts its
same perceived event. (It may be worth noting one impor- net upon the available activated memory (in this case, pre-
tant nuance, that the psychological moment, the time dur- sumably a stream of auditory sensory memory) and picks up
ing which sequentially-presented stimuli are perceived as the 3 to 5 most salient chunks within the representation.
simultaneous, is an order of magnitude smaller; see, for ex- The pre-list silent period makes the sensory information
ample, Allport 1968.) I would not in any way dispute such from the beginning of the list salient and the end-of-list
an interval or its importance in limiting the psychological silent period, as well as perhaps the highly activated state of
present. The stance taken in the target article was just that recent sensory memory, makes that part of the list salient.
these factors could not explain storage limitations in the full In McKones procedure, also, there is the possibility that
range of situations in which they have been observed; for the focus of attention is refilled with the appropriate items,
example, it is hard to see how they could explain the limi- from activated memory outside of the focus, only after the
tations observed by Luck and Vogel (1997). Nevertheless, I recall probe is presented. Thus, time-based forgetting is
agree that there are some likely links between any such psy- only one part of what differentiated the salience of differ-
chological present and the chunk-based capacity limit. For ent items. McKones reaction time function could be ac-
example, a rapid sequence of 4 stimuli presented within a counted for if there were a probabilistic tendency to fill the
1.2-sec window would allow 300 msec per stimulus presen- available 4 or so slots with the most recent items, but if on
tation, just long enough for brief stimuli to be perceived some trials some of the slots were filled with items from
(Massaro 1972; Turvey 1973). Perhaps there is both a tem- other serial positions or not filled with items. The monoto-
poral and a chunk limit to the psychological present, lead- nically increasing advantage across the last 4 serial positions
ing to a rule such as this: The psychological present can in- would occur because, on many trials, the slots were indeed
clude up to 35 chunks of information at once, provided filled with the last 4 items.
that these chunks all can be perceived, conceived from long- McKone also objected that the capacity was much lower
term memory, or reactivated through rehearsal within the than 4, specifically about 2, in the implicit memory of pseu-
period of one psychological present. Such a rule might make dowords. I would suggest here that under implicit memory
many investigators happy (e.g., Baddeley 1986). However, situations, it may not be the case that each pseudoword was
it should be noted that the reason for the rule might not be learned as a single chunk. Instead, many pseudowords may
decay of the first-presented chunks. Rather, chunks pre- be recognized only as several separate chunks, as was sug-
sented farther apart might be difficult to integrate into the gested in the target article (sect. 3.2.1).
same perceptual present (or coherent scene, in present Taatgen presented an ACT-R model that combines a 1-
terms). Given that this rule is stated in such abstract terms, chunk focus of attention with activation, decay, and inter-
considerable empirical work would be needed to assess its ference of information from long-term memory. Except for
validity. It would have to be shown to apply with rehearsal the postulation of only a 1-chunk focus, this model is in
processes blocked. Rehearsal may serve the purpose of al- keeping with that presented by Cowan (1999). On one hand,
lowing the subject to re-present all of the items within one the output of the model (shown in Taatgens Fig. 1) seems
psychological present in order to encode them into a com- very promising. The low-capacity curve appears to match
mon mental context. the data for the unattended speech condition shown in Fig-
Finally, Grondin found the postulation of a chunk limit ure R1, whereas the high-capacity curve appears to match
(rather that a time limit) as central to short-term memory the attended-speech condition and the curve presented by
to be an oxymoron; a term is a time period. Perhaps so, Bachelder. However, it is difficult to understand how the
but the continual need for the focus of attention in pro- function based on the effects of decay, interference, and
cessing practically ensures that any particular set of chunks the increased probability of doing something wrong if more
must leave to make room for other chunks within a rather responses are required (Taatgen) can account for evi-
short term. This explains the great difficulty of vigilance dence with simultaneous arrays followed by only one deci-

sion (Luck & Vogel 1997) unless a metric is proposed for ory is used to respond in the no-change situation, it is use-
mutual interference between elements within a simultane- ful to define performance in terms of the rate of correct re-
ous display. Again, it also is difficult to understand how a 1- jections, CR. The assumptions just stated then lead to the
chunk focus of attention can accomplish comparisons be- following expression: CR 5 k/N 1 [(N 2 k)/N](1 2 g).
tween items. The success of the modeling is encouraging Combining equations, H 1 CR 5 2k/N1 (N 2 k)/N 5 (k
but we probably do not know how many different forms of 1 N)/N. Rearranging terms, the capacity can be estimates
model can accomplish the same thing. I agree with the im- as k5(H 1 CR 2 1)N. The resulting estimates appear to be
plication that capacity, an integer quantity, ultimately is compromised slightly by a performance ceiling in set size 4
likely to result from developmental changes and individual but they are rather constant across larger set sizes, which we
differences in some underlying continuous quantity or have tested up to 10. The resulting capacity estimate in sev-
quantities. However, I see no reason to exclude attention as eral data sets is about 3.5 items.
the basis for that change or the locus of that capacity. To illustrate the working of the model, estimates can be
In sum, I agree with many of the commentators that de- obtained on the basis of the data figure provided by Schnei-
cay (or any type of time-based forgetting) and interference der et al. for a set size of 6. Inasmuch as H1CR equals
are important factors to consider in short-term recall. How- twice the proportion correct P, for large objects or small ob-
ever, the present thesis is that a 4-chunk limit of the infor- jects, assuming P 5 .80 (from the data graph), k 5 (2(.80)
mation in the focus of attention at one time can be identi- 2 1) 6 5 3.6. For combined objects, if we assume that 6
fied apart from these other factors. According to the model small objects plus 6 large objects equal 12 objects total and
of Cowan (1995; 1999), decay and interference determine P 5 .63 (from the data graph), then k 5 (2(.63) 2 1)12 5
what information will be available and accessible to the fo- 3.12. The lower number in the second case suggests that not
cus of attention, from which about 4 chunks will be se- only were large and small objects perceived separately, they
lected. While in the focus, items may be protected from also may have interfered with one another slightly. In sum,
further decay or interference. Evidence presented by Ren- simultaneous arrays lead to quantifiable capacity-limit esti-
sink from change-detection experiments is especially help- mates, much as sequential presentations do.
ful in illustrating this point of the separability of capacity
from other processing limitations. For example, when the
interval between displays increases to 360 msec, eliminat- R3. Cognitive theoretical accounts
ing a visual sensory memory contribution, a change can be of the 4-chunk limit
identified reliably among at most 34 items, and this limit
R3.1. Metatheoretical considerations
does not change with further increases in the interval.
As mentioned above, my strong belief in the 4-chunk limit
is not accompanied by convictions that are as strong re-
R2.4. An improved procedure to estimate
garding the theoretical reason for the limit. Many of the re-
capacity in visual arrays
viewers offered their own theoretical approaches to under-
The majority of the criticisms of the 4-chunk hypothesis standing capacity limits. The theoretical problem we have as
have been directed at the interpretation of verbal, serial re- a field is certainly not the absence of a plausible explanation,
call measures. Therefore, it is important to communicate but rather many different, and in some cases conflicting, ex-
that, since the target article was written, I have developed planations. Therefore, a first step toward the understanding
(with co-investigators J.S. Saults and N.M Fristoe) an im- of the capacity limit may be to set out a simple taxonomy of
proved means to assess capacity in visual array experiments. capacity-limit theories. In doing so I will assume that the
Specifically, we have developed a formula to index memory 4-chunk limit applies. Inasmuch as I already discussed al-
capacity in the procedure of Luck and Vogel (1997) in ternative theoretical approaches that do not lead to the pre-
which a single item within the second array is encircled and diction of a 4-chunk limit in STM, the following discussion
the task is to determine whether that item has changed will be limited to theories that are, at least to my under-
color from its counterpart in the first array (the arrays be- standing, compatible with a 4-chunk limit. The goal is to try
ing otherwise identical). The formula was a modification of to discover what types of experimentation will be most help-
one presented earlier by Pashler (1988). The logic of the ful in distinguishing between these possibilities.
measure is as follows. After making certain core assumptions (from Cowan 1995;
Upon examining a briefly presented array of N items, the 1999), I find that the most general question that arises is
subject is able to apprehend a certain fixed number of items, about the mechanism that constrains capacity. A basic sug-
k. The apprehension of these items would allow a change to gestion in the target article was that the limit is in the ca-
be detected if one of these k items should happen to be the pacity of the focus of attention. Upon reflection, however,
changed item. Thus, with probability k/N, the change is de- one might interpret that suggestion in two different ways.
tected. If the change is not detected, the subject guesses First, the focus of attention might be able to hold only a lim-
yes, there was a change with probability g. Thus, the for- ited number of items from the activated representation con-
mula for the hit rate H is: H 5 k/N 1 [(N 2 k)/N]g. If there currently. Second, the activated representation might be
is no change between the two arrays, and if the cued item limited in some way that constrains how many items can be
happens to be an item that is included within the set k that recovered by the attentional mechanism.
the subject apprehended, then that knowledge will allow These two possibilities are illustrated in Figure R2 for
the subject to answer correctly that no change has occurred processing of a single row of 7 letters in Sperlings (1960)
(and this is where our formula differs from Pashlers). If whole-report task. Before explaining the figure, the under-
there is no such knowledge (for N 2 k items), then the sub- lying assumptions must be stated as they apply to this case.
ject still will answer correctly with a probability 1 2 g, where All 7 items are presumably contained briefly (for several
g is again the probability of guessing yes. Given that mem- hundred milliseconds) in a visual sensory store. This visual

that attention is not a unitary concept and that, therefore,
limits apply separately in different domains. This type of
proposal appears to be a variation on H1 in which different
attentional limits operate at the same time in different do-
mains. There are other possible accounts, but these are ad-
equate for the present exposition.
Some commentators alternative views (sect. R3.2) will
be classified with respect to these four categories along
with further specifications as appropriate. The theoreti-
cal views include separate-stores accounts (R3.2.1), unitary
memory accounts (R3.2.2), multiple-attention accounts
(R3.2.3), and storage-plus-processing (neo-Piagetian) ac-
counts (R3.2.4). Biologically oriented explanations will be
presented and some will be classified according to these
categories later on in section R4. Last, a tentative appraisal
of the models will be offered in section R5.
Figure R2. An illustration of two alternative, cognitive theoreti-
cal hypotheses as to why a capacity limit of about 4 chunks is ob- R3.2. Alternative theoretical views
tained. The illustration is for a 7-item linear character array in a
whole-report condition like that of Sperling (1960) but the depic- R3.2.1. A separate-stores account. Whereas I suggested
tion should be similar for any estimate of pure STM capacity limits. that the focus of attention works with an activated portions
of long-term memory, Baddeley proposed that the focus of
attention works with a number of passive stores. He sug-
sensory story might consist of a feature map that deterio- gested that my characterization of his model contains some
rates unless it is enhanced by attention quickly. (The feature- misinterpretations and pointed out both similarities and
map concept was included because most errors in the differences between our views. A comparison of our ap-
Sperling-type procedure are location errors rather than proaches could be helpful.
omissions or substitutions, as shown for example, by Irwin At the time that I wrote the section on his view of work-
& Yeomans 1986; and by Mewhort et al. 1981.) Each item ing memory, I had not yet heard of his addition of an
is assumed to be encoded as a separate object (or chunk) episodic buffer (and did not realize until now that he con-
because the rapid presentation did not allow enough time sidered it to be of limited capacity in terms of chunks). With
for items to be grouped. Not all of this information is held that addition, I agree that his view and mine are rather sim-
in the focus of attention at one time. Items presumably ilar despite different uses of terminology.
must come into the focus of attention before they can be In the target article I used the word arbitrary to describe
deliberately recalled. In the example, the activated repre- the stores in Baddeleys model, as he noted; but that term was
sentation comes from sensory memory but, presumably, misleading. I meant to imply only that the stores are not ex-
long-term memory retrieval also produces an activated rep- haustive; some types of coding do not appear to fit in well.
resentation. The main difference is that the latter tends to The contribution of the episodic buffer may change this.
have fewer sensory features and more semantic features Baddeley (in press) described this buffer briefly. The main
than activated memory originating in sensory memory. reason for proposing it appears to be that activated elements
Given these assumptions, two loci for the limit in capac- of long-term memory do not include the novel combinations
ity then can be articulated. In the first hypothesis (H1, of these elements that occur in new experiences, which
shown on the left of Fig. R2), the focus of attention, depicted clearly must be preserved if new episodic information is to be
by the dashed circle, is limited to four items within the en- remembered (in both the short and the long term, although
coded set. In the second (H2, on the right), the focus of at- the episodic buffer is for the short term). With this addition,
tention (dashed ellipse) can process all of the items but their the view is either a limited-representation view (H2) or, with
representations interfere with one another. This particular another limit in attention thrown in, a version of H3.
example shows one plausible deficiency, a bow-shaped re- Although Baddeley (in press) contrasted this point to my
call function; items surrounding the center of the display are own view, I actually agree with him. Thus, Cowan (1999, p.
too indistinct from one another within the memory repre- 89) stated,
sentation for them all to be recalled. Variants of both of these Finally, there is one important qualification of the statement
alternative views may be viable, given the available evi- that working memory contains activated elements of long-term
dence. memory. Most stimulus situations in life include novel combi-
Let us refer to these hypotheses shown in Figure R2 as nations of familiar features. In memory the elements are acti-
(H1) the limited-focus view, and (H2) the limited-repre- vated independently, but the particular links between those el-
sentation view. The figure does not capture all theoretical ements are often novel. The current combination of elements
views but it can serve as an anchor point from which other may, however, be stored as new long-term memory trace.
views also can be described: for example (H3). It is possi- Similarly, Cowan (1995, p. 101) previously stated,
ble that there are limits in both the focus of attention and The information that is in an activated state must include new
the representation upon which attention is focused. For ex- information as well as the activated set of previously learned in-
ample, one could propose that the focus of attention can information. Links between items and/or between each item and
clude only one coherent scene at a time and also that the its serial position context must be generated and preserved as
scene can include no more than 4 chunks because of limi- additional activated information. The new links comprise an epi-
tations in the representation. One could also propose (H4) sodic record that will become part of long-term memory.

A remaining, subtle difference between views is that I have Baddeley believes, but I simply have not dealt with any such
envisaged a mechanism different from Baddeleys to ac- further fractionation.
commodate this information. I have imagined that the 3. As mentioned, my focus-of-attention mechanism was
newly combined information immediately forms new paths said to have storage capacity, whereas in Baddeleys recent
in LTM, while at first being in an activated state (and often, conception that capacity resides in the episodic buffer. An
or always, at first in the focus of attention) and then sub- important difference between the predictions of these
siding out of activation. The newly combined information views appears to be that momentary distraction would
includes, for example, the binding of elements to serial po- cause items to be lost from a focus of attention (though the
sitions in the current list. In contrast, Baddeley sees the information would stay in an activated form for a while and
need for a separate episodic buffer to hold the newly therefore could be returned to the focus of attention with-
formed combinations temporarily. out much difficulty). In contrast, distraction would not be
Several questions about the episodic buffer seem to re- expected to have as severe an effect on an episodic buffer.
main unanswered at this point. What passive storage sys- Clearly, the effect of distraction on a limited-capacity sys-
tem would hold, say, imagery for musical timbre (Crowder tem is a fruitful question for future research and may help
1989), which is neither phonological nor spatial in nature? to resolve this issue.
If the episodic buffer holds such a code, does the code then
have a status comparable to phonological and spatial codes R3.2.2. Unitary-memory accounts. At least four commen-
or is the episodic buffer different (e.g., less passive in na- taries (Nairne & Neath; Avons et al.; Rensink; Taatgen)
ture)? Presumably, one can deposit a musical timbre in questioned, in one way or another, a fundamental assump-
memory either passively, through a sensory stimulus, or action underlying the present thesis that there is a 4-chunk ca-
tively, through imagination. A potential concern is that the pacity: the assumption that there exist separate short- and
system, even as amended, seems to suggest that some long-term memory faculties. This assumption is necessary
codes (phonological and spatial) are represented in a more if one wishes to hold the belief that long-term memory is
fundamental or enfranchised manner than others (such as unlimited in capacity, whereas short-term memory is lim-
acoustic quality or, say, tactile quality) and there may not ited. Instead, the theoretical account would presumably
be sufficient empirical support for that particular assump- draw on general principles of memory, such as contextual
tion. cues in retrieval and the distinctiveness of memory repre-
My statement about the focus of attention that no other sentations within a retrieval context. As such, it appears to
mental faculties are capacity limited was followed imme- be a version of H2, a limited-representation view.
diately by a qualifier: although some are limited by time Although this type of account seems reasonable, it seems
and susceptibility to interference. Thus, I use the term ca- theoretically implausible that there is not some sort of STM/
pacity in a narrow sense. I agree that this set of processing LTM distinction. I define STM rather broadly (as the focus
limits is probably similar to Baddeleys (1986) model, except of attention along with other activated elements of LTM;
that chunk limits were not emphasized within that earlier see Cowan 1988; 1995; 1999). It seems to me that some-
description of his model. thing of this nature must exist. First, I think that no one
Baddeley pointed out three apparent differences be- truly believes humans can attend to an unlimited amount of
tween our models, to which I will respond. information at once, and that the concept of an attentional
1. I was said to regard short-term memory as simply the limit is practically self-evident. Second, the concept of mem-
currently activated areas of long-term memory, which was ory activation probably maps onto the distinction between
said to be inconsistent with neuropsychological data inas- information held through synaptic connections (long-term
much as deficits in long-term memory are found which storage) and the subset of that information that is currently
have no apparent impact on immediate memory, or on in an electrochemically active state (short-term storage).
working memory more generally. I cannot claim to be an My view of short-term storage does not require that a sep-
expert on the neuropsychological data. I would caution, arate copy of the information be formed in a different fac-
though, that what appear to be separate short-term stores ulty within the brain, in contrast to a view offered of the ba-
could be, instead, separate routes to the activation or use of sis of neuropsychological considerations by Baddeley;
information in long-term memory, so that it would be spe- though this view was apparently attributed to me by Taat-
cific retrieval processes, rather than stores, that are im- gen, probably because of my use of the term STM to refer
paired or left intact in some patients. Processes versus to this activation, against the wise counsel of Morra. Even
stores can be notoriously difficult to distinguish. though most cognitive researchers may accept the idea of a
2. My limited capacity system was said to be unitary, in limited attention and memory activation of some sort, I sup-
contrast to the fractionable central executive of Badde- pose that many of them do not see these as scientifically
leys model. Actually, I distinguish between the control as- helpful concepts. I would argue, however, that one cannot
pects of attention and the focus of attention (e.g., Cowan apply the general principles without at least an implicit use
1995), in a manner that seems similar to Baddeleys con- of the attention concept to describe capacity limitations
ception. His central executive control processes appear during encoding and retrieval, even if the 4-chunk limit
similar to my central executive or control of attention (heav- turns out to be a result of limitations in the focused-upon
ily involving the frontal lobes), whereas his episodic buffer memory representation.
probably is more closely aligned to what I call the capacity- The unitary-memory theorists presumably want to use
limited focus of attention (heavily involving the parietal general mechanisms of memory to account for the 4-chunk
lobes and adjoining temporal areas, according to sugges- limit. It is worth noting that they have various options open
tions made by Cowan 1995). For the present target article to them. I can identify basically three: overwriting, pertur-
I saw no reason to emphasize control processes. Further bation, and contextual distinctiveness. I will try to describe
fractionation of the central executive may be possible, as the potential relevance and boundaries of each of these.

In an overwriting model (e.g., Nairne 1990; Neath & the context at the time of retrieval. One type of contextual
Nairne 1995) each stimulus is assumed to be represented cue is a timing signal from the time of encoding (Brown et
as a bundle of features, and features of each stimulus are al. 2000). Contextual distinctiveness is the principle under-
said to overwrite the same features that happened to occur lying the ratio rule that has been used to explain the recency
within previous stimuli in the sequence. (By the way, I effect in recall: the principle that the ability to distinguish
would think that long-term memory is preserved intact, two nearby items correctly depends on the ratio of the time
whereas the overwriting takes place in activated memory, in between the items and the time between the last item and
which case this type of model would seem to be misclassi- the recall cue. It could account for the 4-chunk limit in se-
fied and actually is a dual-store model; but its users typically quential recall on the basis of diminishing distinctiveness of
see it as a unitary model.) Murray presented an account of items earlier in the list. An analogy could be drawn so that
this type. Overwriting models of this sort seem possible for spatial distinctiveness could govern the ability to tell apart
the description of memory for sequential stimuli but not si- objects within the visual field. That presumably is the gist
multaneous arrays. It is not clear to me, in this type of of the finding of He et al. (1997) mentioned by Scholl
model, what overwriting is supposed to take place among & Xu, showing that with great enough spatial distinctive-
simultaneously presented stimuli as in the Luck and Vogel ness one can enumerate numbers above 4 in a visual after-
(1997) task in a manner that depends on the number of image; although I have no access to the research report.
stimuli in the array. Thus, someone with this view might The theory of subitizing mentioned by Avons et al., based
well be forced to suggest that capacity limits in simultane- on the application of the Weber fraction for a continuous
ous arrays arise from a different mechanism than memory perception of numbers, may be an example of the distinc-
limits in sequential presentations. Then it is not clear if the tiveness principle but based on a perceptual representation
limit of about 4 chunks in each realm would be judged the (and sometimes its afterimage) rather than a longer-lasting
result of a coincidence or some deeper design principle. In memory representation. It involves the comparison of the
any case, if one accepts that attention must be focused on perceived quantity with a continuum of quantities encoded
the representation in order for recall to take place, it is a in long term memory. The alternative account of Peterson
version of H2 or H3. (The same probably could be said for and Simon, mentioned by Taatgen, can be described sim-
perturbation and contextual distinctiveness; all of them fo- ilarly except that it is a pattern rather than an amount that
cus on limitations in the representation rather than in at- is compared with the contents of long-term memory.
tention per se.) There are other principles of memory but they may not
In a perturbation model (e.g., Lee & Estes 1981; Nairne apply directly to the question of the reason for the chunk
1991), the items do not obliterate one another in memory limit in memory. For example, memory can be improved
as in the overwriting model. Instead, representations of with long-term recoding in order to form associations be-
items are said to switch serial positions within the memory tween items and between these items and existing struc-
representation of a list, a process termed perturbation. tures in memory. However, by definition, this type of
Usually, it appears to be nearby positions that switch. One process alters the size of chunks rather limiting how many
could also generalize this model to a spatial array with the chunks can be recalled, and so is not relevant here. Now, let
assumption that adjacent spatial positions switch within the us turn to some of the evidence provided by commentators
representation. This type of model widens the types of rep- taking a unitary-memory view.
resentation that can be considered. Whereas an overwrit- Nairne & Neath presented a provocative result. Words
ing model may be inherently committed to some sort of presented in lists for pleasantness ratings were presented
temporary representation because an overwritten trace is again in alphabetical order, the task being reconstruction of
destroyed, one can talk of perturbations taking place in a order in the long term. Lists of an average of 3.75 items
representation without it being destroyed. Nevertheless, if could be ordered correctly 50% of the time. Their conclu-
one holds that long-term memory representations are not sion was, reasonably enough, that a limited capacity does
destroyed, it is a bit mystifying how they can be so altered. not indicate a separate STM mechanism.
Perhaps the perturbations take place at the time of re- This experiment tells us a great deal about the mecha-
trieval. If so, multiple retrievals from the same list might nisms of working memory, which will be addressed in later
tend to show a perturbation during one retrieval that is un- sections. However, it does not pose a problem for the pres-
perturbed in a later retrieval. It is not clear to me how good ent theoretical view. As Shiffrin (1973) noted: in a sense,
the evidence for a perturbation mechanism is because every deliberate task is a short-term memory task. The mem-
guessing factors in serial recall can produce patterns of re- ory representation must be brought to consciousness before
sults that look like perturbation patterns (Anderson & Ma- it can be overtly recalled. Thus, it is perfectly possible for the
tessa 1997). On the other hand, the target article brought representation of a list to be retrieved from long-term mem-
up the possibility that items that had been encoded in the ory into an activated state so that the most distinctive aspects
same attention cohort tend to be confused with one another of the representation (limited to about 4 at a time) can be
(sect. 3.4.3) and that could be the mechanism of perturba- drawn into the focus of attention and overtly recalled.
tion (which would seem to place perturbation in the camp If there are separate STM and LTM faculties, the STM
of H1 rather than H2 or H3, where it otherwise would fall). faculty could affect LTM recall not only at the time of re-
Finally, in a contextual distinctiveness model, the sets of trieval as Shiffrin noted, but also at the time of encoding.
features established during encoding are not necessarily With the latter case apparently in mind, Broadbent (1971)
degraded but there is a failure to distinguish some items (or addressed the STM/LTM relationship as follows (pp. 342
chunks) from others within the representation at the time 43):
of retrieval, and therefore an inability to recall all of the There remain to be considered two points urged by interfer-
chunks appropriately. This confusion among items can de- ence theory: the existence of effects on short-term memory
pend on the similarity between the context of encoding and from previous long-term experiences, and the continuity which

seems to exist between memory at long and short periods of theoretical treatments of the phenomena mentioned. There
time. The first of these must be admitted straight away, and is was no discussion of the distinction between controlled and
perfectly consistent with a view of short-term memory as due automatic processing in visual search and memory search
to recirculation into and out of a decaying buffer storage . . . In (Schneider & Shiffrin 1977; Shiffrin & Schneider 1977).
general one must beware of concluding that the appearance in The experimental procedure in which search was investi-
short-term memory of an effect known from longer-term stud-
ies is evidence for identity of the two situations . . . Only the
gated during a memory load (Woodman et al., under re-
success or failure of attempts to show differences between the view) was one in which the search targets were the same
two situations is of interest in distinguishing the theories. from trial to trial and therefore was presumably an auto-
matic search. It would be expected that, in contrast, a con-
Pothos & Juola reinforced the logical case introduced by trolled visual search would be influenced by working mem-
Broadbent in the previous quote, that STM tasks and LTM ory load. There also was no mention of the alternative views
tasks would not be expected to be process-pure. He noted on the reason why a response bottleneck often is obtained
that If language is at least partly learned, linguistic depen- (Cowan 1995; Meyer & Kieras 1997).
dency structure should reflect properties of the cognitive Woodman et al. are on to something but not, I believe,
components mediating learning; one such component is something relevant to the 4-chunk hypothesis. It seems rea-
STM. He showed higher contingencies between words sonable to believe that mechanisms contributing to auto-
within a range of 4 words in natural language, with remark- matic visual search and to response bottlenecks occur, re-
able convergence among samples from different languages. spectively, before and after a central attentional process
Like Nairne & Neath, another commentary, Avons et that is linked to subjective awareness. These input- and
al., similarly argued for a continuity of STM and LTM. It is output-related processes still may be termed attention if
precisely this sort of thrust that led me to comment in the one prefers, but their limitations may not be appropriately
target article that the existence of memory decay (i.e., time defined in terms of a capacity limit. They may be limited in-
limitations in certain kinds of mental representations) still stead by other factors, such as the time of processing and
has not been established after all these years. However, do the presence of domain-specific or feature-specific inter-
Avons et al. actually doubt that the focus of attention can en- ference. Therefore, I do not think that they alter the argu-
compass only a small amount of information, in contrast to ments of the target article, although they clearly contribute
the vast amount of information in memory? I believe that this a great deal to our overall understanding of information
dual process, which was central to the target article, should processing.
be counted as a discontinuity between STM and LTM. Scholl & Xu argued more directly that the phenomena
Taatgen suggested that short-term memory capacity is discussed in the target article dealing with visual informa-
not something that can be used to explain the outcomes of tion processing come from a different source than the phe-
experiments, but is rather something that needs to be ex- nomena dealing with verbal information processing. For ex-
plained itself. Why shouldnt the arrow of causality go in ample, they said that a general STM-based theory capacity
both directions? Just as gravity helps to explain many phe- limits could not easily account for the strong dependency
nomena (e.g., our ability to stay on earth) and a theory of of MOT (multi-object tracking) performance on subtle vi-
space/time attempts to explain gravity, STM capacity may sual details such as the type of accretion and deletion be-
help to explain many phenomena (e.g., problem-solving hind occluders. I do not understand why this is the case.
proficiency) and yet itself needs to be explained at a finer- The findings appear to indicate that tracking performance
grained level of analysis. is preserved with a logical disappearance and reappearance
To sum up, unitary memory accounts argue against a sep- of objects (disappearance behind occluding objects or even
arate short-term storage facility. However, it is unclear virtual objects) but is impaired when objects inexplicably
where they stand on the existence of a limited-capacity at- disappear and reappear. These studies point to a fascinat-
tentional mechanism or on some definitions of memory ac- ing perceptual apparatus that influences the input to a cen-
tivation. (Clearly, even in these unitary views, there is a spe- tral capacity-limited process. If one object disappears and
cial accessible status for information in memory that has not reappears somewhere else inexplicably, it does not pick up
been overwritten and that matches currently available re- the same object identity. However, the capacity limit still
trieval cues.) In a dual-storage view, a continuity between could reside in a central attentional faculty fed by this per-
short- and long-term memory procedures is nevertheless to ceptual process, not in a separate visual attention mecha-
be expected, on the basis of the involvement of these at- nism per se. The ability to track 4 targets in the MOT task
tentional and activation factors both at the time of long- and simultaneously acquire and hold 4 verbally-presented
term learning and at the time of retrieval (Broadbent 1971; items in STM was used as an argument against a central
Shiffrin 1993). capacity limit inasmuch as attention-switching is impossible
in this procedure. However, in adults, an articulatory loop
R3.2.3. Multiple-attention accounts. Several commenta- (Baddeley 1986) would be able to hold 4 items almost ef-
tors (McKone; Rensink; Scholl & Xu; Woodman et al.) fortlessly, leaving the attentional mechanism free for track-
argued that the attention-based account is not suitable being. If verbal memory were loaded to capacity or the pro-
cause attention is not a unitary phenomenon. These are ap- cedure were used in children old enough to rehearse, but
parently versions of H4 above. not effortlessly (Guttentag 1984), I would predict more in-
Woodman et al. discussed how visual search occurs on terference between tasks.
one time scale and working-memory search on another The other arguments also seem questionable. Arguments
time scale, and how visual search can take place without based on the subitizing procedure are open to question
impairment when working memory is loaded. They also given the uncertainly as to the cause of that phenomenon,
mentioned research on the separability of response limits. acknowledged above (sects. R2.3.3 and R3.2.2), although it
I believe that their arguments overlook some important may well be that common principles underlie subitizing and

memory limits (much as Newtons inverse-square law ap- The reason that I still do not agree with the limit of 7
plies not only to gravity, but also to other phenomena such is that the views leading to that estimate by Morra and
as light intensity). If subitizing limits come from limits in Pascual-Leone do not appear to allow that processes can
long-term recognition, it stands to reason that uncrowded become automatized and thereafter cease to take up ca-
sets of objects could be more likely to be recognized than pacity. The research combining processing with storage
crowded sets. Arguments based on neurological dissocia- loads (reviewed, for example, by Baddeley 1986) does not
tions are open to question because a deficit in one area (e.g., seem to suggest that there is a strong tradeoff between the
verbal STM) with a normal performance in another area two. In many studies, a moderate storage load proves to
(e.g., subitizing) could be explained on the basis of a defect have minimal effects on concurrent processing. A full stor-
in a perceptual or representational process feeding into the age load has notable effects on processing but, in that situ-
central attentional faculty, rather than a fractionation of the ation, one can argue that the central executive has become
attentional faculty per se. The points raised by Scholl and involved in special processes (various types of complex,
Xu suggest some important possibilities for future research attention-demanding rehearsal) in order to manage the full
to resolve these issues. storage load. Given the disagreement even among the neo-
Some points make by McKone and Rensink can be Piagetians, it is clear that more empirical work is needed to
thought of similarly. McKone noted that Clear theoretical determine the relation between processing and storage. I
reasons exist for limits on the number of items that can be believe that the procedures used in the target article to in-
simultaneously active within a single domain, since leftover dicate a storage limit of 4 chunks are procedures in which
activation from previous items will interfere with identifi- everything except storage tends to become automatized
cation of the currently-presented item. Here one must dis- and takes up little capacity.
tinguish between similarity-dependent interference, on
one hand, and a true capacity limit, on the other hand. Both
may well occur. Theoretically at least, the difference would R4. Biologically oriented accounts
be that a capacity limit depends on the number of chunks of the 4-chunk limit
in memory at the same time, rather than specific similari-
ties in the features of those chunks; whereas interference The present section complements the previous one by ex-
depends on similarity rather than number per se. Rensink amining commentaries describing theory and evidence rel-
pointed to evidence for at least 10 items probably re- evant to the neurobiological underpinnings of the capacity
sulting from the grouping of items of similar contrast sign. limit. These include psychophysiological evidence (sect.
No doubt such a visual mechanism (one that initially breaks R4.1), defined broadly to include electrophysiology and
the world into parts) occurs, but it is presumed to be neuroimaging; evidence from individual and group differ-
a preattentive mechanism that feeds into the capacity- ences in humans (sect. R4.2); and evidence from animal be-
limited process. This type of distinction is reminiscent of havior (sect. R4.3), which is relevant to evolutionary con-
the distinction (Norman & Bobrow 1975) between data- siderations. Following these topics I will discuss several
limited processes, such as specific processes in visual per- neurobiological models of capacity that were offered by
ception, and resource-limited processing, such as central commentators (sect. R4.4) and will seek clarification of the
attention, a distinction that seems very important. Again, I most useful relations between theoretical models and em-
maintain that data-limited processes are affected by decay pirical evidence, and between neurobiology and behavior,
and interference factors but are not truly capacity-limited, in understanding capacity limit (sect. R4.5).
unlike resource-limited processes. Nevertheless, I am sym-
pathetic to Rensinks conclusion that Given that there are
R4.1. Psychophysiological evidence
no compelling a priori grounds which can be appealed to,
this matter will have to be settled by experiment. The evidence that was presented offers some important op-
portunities for further research in the near future. Rypma
R3.2.4. Storage-plus-processing accounts. Three of the & Gabrieli showed that the brain response to one or three
commentaries are from a neo-Piagetian perspective (Hal- items is far less than the brain response to six items. This
ford et al.; Morra; Pascual-Leone). What all of these ac- finding may offer the opportunity to use the neuroimaging
counts have in common is that they suggest that processes to help solve the difficult problem of how to define chunks,
take up capacity. The processing accounts all incorporate discussed in section R2.2. For example, if one presents a
the need for storage in one way or another but that is a ba- sequence such as F-B-I-C-I-A, will the brain metabo-
sis of controversy. The accounts nevertheless all appear to lism look like two simple items (assuming that the subjects
be varieties of H1, suggesting that a central attentional limit form two higher-order chunks) or like six? If it looks like
explains the processing and storage limitations. Halford et six, could it be made to look like two by presenting pre-
al. focus on processes, suggesting that humans are limited exposures of the acronyms? If the method does not match
to relating four entities, (see also Davis), whereas Morra behavioral results, such a discrepancy could suggest that
and Pascual-Leone consider storage and processing sepa- there are hidden processes that need to be added to the
rately and suggest that the total capacity is therefore 7 cognitive model. This type of research could help to answer
chunks (some taken up by processing schemes), not 4. Iron- Tiitinens plea for research to address the chunking issue.
ically, the latter two theorists (like Raffone et al.) appear Gratton et al. considered a distributed capacity hypoth-
to fall into my camp in considering storage and processing esis in which a limit of about 4 chunks would result from a
to share a common resource, as opposed to the suggestion limit of about 2 in each hemisphere. They showed interest-
by Halford et al. that processing limits cannot be sub- ing evidence that reaction times were about the same when
sumed under storage limits. Yet, it is Halford et al. who five items were distributed to the two hemispheres 23 or
come up with a capacity estimate in keeping with my result. 14, but not when all five items were presented to one hemi-

sphere. It would seem, though, that a distributed capacity to look for subnormality in schizophrenics given that they
model should predict a smoother transition between the two do not differ in some of the temporal aspects of STM be-
conditions. The result that was obtained may be more con- havior. In particular, they are abnormal in the precision of
sistent with a model in which groupings play a role. Groups auditory encoding, but not in auditory recognition speed or
of 1 to 4 items can be chunked together. Perhaps two such auditory memory duration when the overall performance
chunks can be accessed and searched in parallel; quickly, levels are equated with normal individuals (Javitt et al.
given that they are small sets. When all 5 items occur in one 1997; March 1999).
location, it may be more difficult to separate them into sub- It is not only subnormal individuals who are potentially
groups, in which case a slower search of set size 5 has to be instructive. As discussed in section R2.3.3, Wilding illus-
used. This hypothesis could be investigated by dividing the trated that capacity limits of about 4 items emerge in the
stimuli into two groups above and below fixation, rather form of the number of items clustered together, within spe-
than to the left and right of fixation. The distributed mem- cial subjects whose memory for digits is supernormal. The
ory model should predict no difference between distribu- pattern is similar to individuals who have trained them-
tion conditions, whereas a grouping account predicts a dif- selves to have a supernormal digit span, discussed by Erics-
ference comparable to what was obtained in Gratton et al.s son & Kirk. In both cases, there are anomalies that the au-
Figure 1. This, too, is a promising area for research. thors would want to attribute to some mechanism other
Jensen & Lisman brought up several new physiologi- than a capacity limited to 4 chunks of information, though
cal findings related to the hypothesis of Lisman and Idiart in both cases the predominant finding is consistent with the
(1995) that nested cycles of oscillation in the brain can ac- 4-chunk hypothesis. If capacity limits are eventually to be
count for the capacity limit. These, too, show promise al- understood well, these anomalies in the results must be un-
though, sooner or later, a tighter linkage between physio- derstood; but it still seems to me that the vast preponder-
logical and behavioral results will be needed (as discussed ance of evidence is favorable to the 4-chunk hypothesis. We
in sect. R4.5). For example, how can the physiological ev- all would agree that these individuals unusual abilities can-
idence for changes in oscillation rates in the brain be rec- not be attributed to a larger-than-normal memory capacity,
onciled, if necessary, with behavioral evidence favoring a and therefore that supernormal performance depends
parallel rather than serial memory-search process, as dis- more on training than on biology. Yet, it remains to be de-
cussed in section R2.3.3? It is helpful that several investi- termined if any normal individual could learn to accomplish
gators have begun to find behavioral evidence of changes these mnemonic feats. Given that pure STM capacity mea-
in stimulus presentation rates that can be closely linked to sures have not been applied, we cannot rule out the possi-
the physiological data on the special role of the 40-Hz bility that there is a minimum level of capacity that an indi-
rhythm in humans; not only Burle and Bonnet (2000), as vidual must have before being capable of becoming a
brought up by Jensen and Lisman, but also Elliott and mnemonist. It must be difficult to distinguish between raw
Mller (1998; 2000). Clearly, though, it will be a while be- ability and motivation, with some individuals falling by the
fore we know what is happening in these innovative new way instead of becoming experts because of insufficient
procedures. motivation or inadequate ability.
Hecht & Shackelford remarked upon the possible role
of pure STM capacity limits for applied areas of cognitive
R4.2. Evidence from individual and group
development, including mathematics. Of course, this gen-
differences in humans
eral approach is consistent with the neo-Piagetian ap-
As discussed in the target article (e.g., sect. 3.1.3), studies proaches of Halford et al., Morra, and Pascual-Leone.
of individual and group differences can contribute to our The assumption is that biological changes mediate a growth
understanding of the mechanisms underlying capacity lim- of capacity developmentally. Hecht and Shackelford ap-
its. Insofar as the individual and group differences in ca- pealed to Gearys (1993; 1995) distinction between biolog-
pacity can be attributed to biological sources, they have the ically primary skills, which come relatively easily in devel-
potential to yield information about relevant biological pro- opment (e.g., natural language learning; math in the
cesses. Oades & Jemel considered the possibility that the subitizing range), and secondary skills which are much
capacity limit may be different in individuals with schizo- newer in the evolution of the human species and prove dif-
phrenia. It is a good point that pathological populations are ficult for many children (e.g., learning to read; math in the
a likely source of information about the reason for capacity counting range). According to Gearys theory, the secondary
limits. Oades and Jemel brought up an interesting question. skills can be learned through a co-opting process in which
Given that schizophrenics are renowned for loose associa- primary skills are used to bootstrap the secondary ones.
tions, why do they not chunk items more than normal indi- Hecht and Shackelford wrote that A limited pure short-
viduals and therefore show a supernormal STM? I think term memory capacity, shaped by evolutionary forces, may
that the answer must be that the loose associations are non- currently be co-opted for many contemporary tasks such
selective and are based on characteristics of the stimuli that as biologically secondary mathematical problem solving
are typically irrelevant for performance. Perhaps if those skills. If this is the case, measures of pure STM capacity
typically irrelevant associations were make relevant, schiz- could be important in understanding individual differences
ophrenics would excel. Oades and Jemel stated that schiz- in applied areas. One reason why these measures could
ophrenics excel on global processing. They might excel also serve a different purpose from more complex measures of
if tested on, say, incidental occurrences of rhymes among working memory (e.g., Daneman & Carpenter 1980; Dane-
stimuli to be memorized on a semantic, rather than a man & Merikle 1996) is that the latter can incorporate many
phonological, basis. In this case, their loose associations processes (item storage, sentence or task information pro-
could result in the recoding of relevant stimuli into fewer cessing, and probably rehearsal) and therefore do not make
chunks. Pure capacity limits are an important area in which it clear which basic abilities underlie the correlation of

working memory with practical skills. Measures of pure cal in humans, chimpanzees might have a pure capacity of
STM capacity might be questioned on similar grounds but about 2.5 chunks on the average. That would make them
are at least more narrowly focused than complex measures similar to the first-grade children studied by Cowan et al.
of working memory capacity. (1999) in an unattended-speech procedure, consistent with
the claim of Kawai and Matsuzawa (2000) that their chim-
panzees performance is the same as (or even more than)
R4.3. Evidence from animal behavior
preschool children. Alternatively, it is possible that adult
Several interesting biological questions can be addressed chimpanzees have a pure capacity limit of about 4 chunks,
through comparisons of animals to humans. First, in a pure similar to humans, and that the absence of sophisticated
STM capacity measure, if capacity is linked to uniquely hu- strategies means that there is less of a difference between
man abilities, there should be a substantial advantage for pure and compound limits than there is in humans. Conse-
humans. The absence of any advantage for humans would quently, this study raises important questions that can only
indicate that capacity is biologically more fundamental. be answered in future research.
Second, in a compound STM measure, if performance de- Todt presented evidence that songbirds repeat songs in
pends on skills unique to humans, there should again be a packages of about 4 chirps, except when the songs are
substantial advantage for humans. (It is reasonable to as- spaced out over time, in which case performance is reduced
sume, for example, that verbal rehearsal strategies are below that chunk size. It is unclear if the distinction be-
unique to humans and assist their performance.) The ab- tween pure and compound capacity estimates is a mean-
sence of any advantage for humans would suggest that any ingful one in songbirds or not. If one can assume that this
strategies unique to humans are not assisting performance. is a pure capacity estimate in the songbirds, it suggests that
Kawai & Matsuzawa described the interesting research the basic capacity limit is a fundamental property of effi-
from their recent paper in Nature, implementing a type of cient brains and not a signature of higher human intellect.
memory span task in a numerically trained chimpanzee, in It would then be a faculty that is, in the terminology of
which digits appeared in a random array about the com- Geary (1993; 1995), co-opted for human intelligence;
puter screen and were to be selected by the chimp in nu- useful for many things including holding information in
merical order, the digits identities being covered by masks mind while processing it in a sophisticated manner. The
on the screen after the first response was made. A reason- similarity between evidence on the burst size in songbirds
able argument was made that the chimps remembered 5 and the cluster size in human recall (e.g., sects. 3.4.1 and
digits 65% of the time. (The first digit in the sequence of 5 3.4.2) is striking. The mystery is only how it could come
could have been identified visually, whereas the remaining about that estimates of capacity in human children (e.g.,
digits were masked; but the response lag was relatively long Cowan et al. 1999) fall below estimates in songbirds. That
before the first digit and was shorter between digits, sug- could be the result of an inappropriate comparison between
gesting that the first digit was part of the memorized set.) different methods. Alternatively, it could be an authentic
Five adult human comparison subjects were run. Their pro- phenomenon, analogous to the observation that motor de-
portions correct were not reported in the article but their velopment is in many ways more advanced in adult animals
mean accuracy in the same condition as the chimpanzee than in immature humans.
was 97% with 4 numbers, 92% with 5 numbers, and 83%
with 6 numbers (Kawai, personal communication, August,
R4.4. Neurophysiological models of capacity
2000). This suggests that adult human spans are about 2
items higher than the chimpanzee. Three commentaries briefly described explicit neurophysi-
Arguments could be found for this task being either a ological models that can predict capacity limits of about 4
pure STM capacity measure or a compound STM measure. chunks (Jensen & Lisman; Raffone et al.; Usher et al.).
I suspect the latter. The stimuli appeared in a simultaneous Probably the most important lesson here is that there are
spatial array on the computer screen, as in many pure STM multiple ways to accomplish the same thing. I must admit
capacity measures. In this case, however, the presentation straight off that I do not profess expertise in the workings
time was not short and was determined by the subject. The of these models, severely limiting how much I can say about
stimuli lent themselves to coding as a coherent series in a them.
particular spatial configurations, and therefore might not Jensen & Lisman elaborated upon the type of oscilla-
have been too taxing to bring from sensory memory into a tory model that was described by Lisman and Idiart (1995)
categorically coded form. When humans participated in a and discussed in the target article (sect. 4.1.2). Rather than
second condition in which the numbers were masked after adhering strictly to Millers magical number, they state on
a fixed 750 msec period from the beginning of each trial, the basis of recent empirical evidence constraining their
their scores were much more similar to those of the chim- modeling parameters, that the upper limit of STM is about
panzee (Kawai, personal communication, August, 2000). 5 to 6 items. My main question regarding this model is how
Attention to the stimuli at the time of their presentation absolutely it depends upon the memory search process be-
apparently can assist performance, not only in humans ing serial in nature. The model is of obvious heuristic value
through verbal rehearsal but also through a mechanism but ultimately it will be important to see how it could, for ex-
available to chimpanzees. (This should be clear also from ample, produce recency advantages within search, and how
the benefit of attention during presentation in first-grade plausible the additional necessary assumptions would be.
children who should be a bit young for strategic rehearsal Raffone et al. presented a model that appears to be less
[Cowan et al. 1999].) Thus, the study does not provide an wedded to the notion of a serial search process. In their
estimate of the chimpanzees pure-capacity limit. model, neural assemblies in high-level visual areas, coding
One might speculate that if the pure limit in chim- unrelated features or objects, exert mutual inhibitory or de-
panzees is about half of the compound estimate, as is typi- synchronizing actions. By stating that competition and

desynchronizing actions between neural assemblies coding oretical statement may not be right, it does have empirical
for unrelated features are stronger within than between consequences that differ from other theoretical concep-
specific representational domains, they appeared to be ad- tions (as shown in Fig. R2 and accompanying text). Milner
mitting both representation-specific and general types of also objected that the neural substrate of attention is not a
limitations (Hypothesis H3 above). One important ques- unitary system that can be pointed like a spotlight or a cam-
tions about this model is what the role of attention may be era, much less a static process into which peripatetic images
(for the general limitations). The conclusion stated in the can be directed. Apparently it must be a highly organized
commentary was that only four magical neural assemblies system of centrifugal paths, every bit as specific as the cen-
can dominate the brain at any given time, saving other tripetal sensory paths that it modulates. I tend to agree but
domain-specific active neural representations through se- would urge caution when mixing levels of analysis. A cen-
lective synchronization. Within this scheme, what happens tral circle for the focus of attention in my cognitive model
to unattended information that is active in memory? My (Cowan 1988; 1995) was not intended to imply geographic
own view (e.g., Cowan 1999) has been that the limit of four unity of the focus of attention in the brain. It instead was
neural assemblies applies only to attended information and meant, at most, to refer to a more abstract kind of unity,
that, outside of attention, numerous other features can re- where all elements in the focus of attention adhere to a
main active in memory (although not fully assembled like common theme (i.e., scene coherence). Moreover, the uni-
the attended chunks). It is not clear if this model fully tary nature of the representation of activated information
agrees with my view or not and, if not, whose conception is (surrounding the focus of attention) is no more than an
correct. oversimplification in the drawing, inasmuch as the theoret-
Last, Usher et al. similarly presented a mode designed ical view places or limits on requirement of coherence on
to overcome shortcomings of the Lisman and Idiart (1995) what can be activated at any moment.
model. It seems similar to the model of Raffone et al., at Roelfsema & Lamme concluded that, the clarity with
least in its orientation and intent (with a reliance on the con- which the psychophysical data point to the number four is
cept of competition between neural networks). An impres- not matched by a similarly clear limit imposed by physio-
sive aspect of this model is that it presents another possi- logical mechanisms. That sort of consideration may per-
bility that did not emerge from my analysis of the behavioral haps provide a motivation to increase the amount of com-
literature. It was stated that an inhibition parameter is set plexity in neural modeling of the capacity limit, as Raffone
high when selection is required and moderate when multi- et al. and Usher et al. did. It also may be worth noting the
ple items need to be maintained together, as in immediate obvious point that, in attempting to find the correspon-
recall tasks. Yet, the capacity of the system is limited to 3 dence between physiology and behavior, both of them have
5 items and cannot increase its capacity beyond this range to be examined very carefully. Roelfsema & Lamme distin-
by diminishing the inhibition parameter even further. The guished between base groupings (combinations of simple
limit was rationalized by the observation that one cannot al- features) and incremental groupings (combinations of
low less representational overlap (in order to increase the more abstract features) and proposed that only incremen-
limit) because representational overlap is essential for tal groupings are capacity-limited, whereas base groupings
computations that perform generalization. This model ap- are rapidly available with no clear limit on the number
pears to be a version of H1 above (a single attentional limit), of base groupings that can be computed in parallel during
but with a flexible attentional focus ranging from 1 to 5 stimulus presentation. For example, it was stated that the
chunks depending on the intensity of focus. What I charac- activation of a neuron tuned to, for example, red and verti-
terized as the singular coherent scene could be the result of cal provides a base grouping between these features. This
a singular focus of attention when it is in the intensely fo- distinction seems to leave open a question about how an in-
cused mode. As evidence continues to accrue, it should dividual would find a red vertical line among red horizon-
have implications for which model is most promising. The tal lines and blue vertical lines. We know that this requires
model of Usher et al. appears to be a version of H3 because a slow, capacity-limited search process (Treisman & Gelade
the capacity depends partly on a concept intrinsic to the 1980) but it is unclear if Roelfsema & Lammes conception
memory representations (the amount of overlap) and partly classifies this as a situation handled by base or incremental
on a concept related to attention (level of inhibition). How- groupings.
ever, it is an interactive concept because neither the repre- Tiitinen asked what kind of research could best link neu-
sentation alone nor the level of inhibition alone state a ca- rophysiology to behavior. He presented some objections
pacity; that is stated by the interaction of these parameters. to the idea that 40-Hz oscillations per se must mediate a
Such a model might be usefully tested in relation to pa- limited-capacity store, given that these oscillations occur in
tients with frontal lobe damage. According to a reasonable an attenuated form even in control conditions. I think that
extension of the model, the ability to control the focus of at- part of the answer to this challenge is to proceed with re-
tention might be altered (e.g., making a narrow focus im- search examining behavioral consequences of neurophysi-
possible) while the maximal capacity limit (35 chunks) ological concepts such as 40-Hz oscillations (already begun
would not necessarily be affected, depending as it does on by Burle & Bonnet 2000; Elliott & Mller 1998; 2000). An-
the degree of representational overlap. other part is to conduct neurobiological research to capture
Several other commentators (Milner; Roelfsema & patterns that have been observed only behaviorally. An ex-
Lamme; Tiitinen) did not fully present models of capac- ample is the finding of Rypma & Gabrieli that the brain
ity limits, but used modeling concepts to challenge or query reaction is much different for 1 or 3 items versus 6 items,
aspects of the present approach. Milner asked, how much and the suggestion of using that finding to explore the way
does it further our knowledge to be told that the span is lim- in which items are chunked together. All of this brings up
ited because the focus of attention can hold only about 4 another issue, which is what neurobiology and behavior,
items at a time? I have tried to show that although this the- and what evidence and models, have to offer one another.

R4.5. Relation between evidence and models sponse. This type of research can help in several ways. If it
in cognition and neurobiology produces brain response patterns that are in accord with
cognitive theory (e.g., the different responses for arrays of
We have seen that there is considerable ongoing work re- sub- versus supra-capacity numbers of items; see Rypma
lated to capacity limits. It is of both an empirical and a & Gabrieli), that then strengthens the cognitive conclu-
theoretical nature, and is occurring in both cognitive and sions by showing an orderly and potentially interpretable
neurobiological fields. (Cognitive research on popula- pattern across levels of analysis.
tions other than adult humans can be considered in this Any discrepancy between cognitive and neurobiological
context to encroach upon neurobiological topics, inasmuch results, indicates that something unexpected may be going
as the differences in brain structure between types of in- on behind the scenes. In that case, however, the discrep-
telligent organism should count as neurobiological quasi- ancy may be difficult to interpret without the help of neu-
manipulations.) In this section, I would like to offer a few robiological theory. For example, the complaint of Roelf-
musings about where all of this research may lead, rather sema & Lamme that neural populations do not tend to
than leaving everything hanging. It is important to consider form the oscillatory patterns needed for a model such as
the logical relations between four types of research en- that of Lisman and Idiart (1995) may provide a motivation
deavor: (1) empirical research on cognition; (2) empirical to look for more complex theories, such as those of Raffone
research on neurobiology; (3) theoretical modeling of cog- et al. and Usher et al.
nition; and (4) theoretical modeling of neurobiology. If the Finally, neurobiological theory, spurred on by neurobio-
aim is to account for a cognitive phenomenon, the pure logical as well as cognitive evidence, can contribute to cog-
STM capacity limit, the relation between these endeavors nitive theory. It can do this by serving as the implementa-
is in some cases asymmetrical. tion of the cognitive theory at a finer-grained lever of
Some cognitive theory is necessary even to observe a cog- analysis (Marr 1982). It also can help by leading to possi-
nitive generalization. Thus, the observation that only 4 bilities that otherwise would not have occurred to the cog-
chunks can be held in working memory could be made only nitive investigators. For example, the use of representa-
with some assumptions about what counts as a chunk. How- tional overlap and inhibition concepts (Usher et al.) led to
ever, the simplicity of the answer that resulted lends some the understanding that the same focus of attention could be
support to those theoretical assumptions by allowing the limited to as few as one chunk or as many at three to five,
detection of a simple pattern in the data (consistent with depending on the setting of an inhibition parameter.
Occams razor). To go further, a theory of capacity must be
stated in cognitive terms so that the elements in the theory
lead to predictions in cognitive terms. Figure R2 summa- R5. Appraisal of models
rizes some attempts along those lines and necessarily intro-
duces abstract concepts (active memory representations; an In this final section I will briefly try to extract, from the im-
attentional focus) that can be directly related to behavior. mense amount of information provided by the commenta-
Neurobiological research is of two types. In some cir- tors, an improved understanding of the basis of capacity
cumstances, manipulations of neurobiological factors cause limits. I will do this with respect to the models illustrated in
cognitive behavioral differences, and a causal link can be Figure R2 and accompanying text.
drawn. We have not reviewed any evidence of exactly that The main question I will address is whether the 3- to 5-
type but it is possible in the future. For example, it is rea- chunk capacity limit in adult humans is likely to be caused
sonable to believe that alcohol may affect working memory by the nature of focus of attention (H1), the nature of mem-
by altering the capacity of attention (see Steele & Josephs ory representations (H2), or both (H3). It seems obvious
1990). However, a neurobiological manipulation has one from the literature that was reviewed in the target article
limitation that is the same as for a cognitive manipulation: that attention exists and that performance can be limited
in either case, the manipulation may not be very specific in similarly either by removing attention or by limiting its ef-
its effect. Any manipulation, biological or behavioral in na- fects (e.g., through articulatory suppression or presentation
ture, that affects capacity only by affecting every single cog- of a complex array of items); so I will not entertain H4. The
nitive function probably reveals little about the mechanism; question for me is just where the pure STM capacity limit
more specific effects are more instructive. On the other comes from.
hand, a large neurobiological change that results in little Several types of study that do not yet exist seem rather
change in capacity limits is potentially very instructive. For crucial (and, indeed, several laboratories are working on
example, if it turns out that lower animals have a capacity them). First, we need studies using two concurrent pure-
limit that is the same as in adult humans (see Todt), that capacity tasks, in which memory loads in very different do-
would show a dissociation between the capacity limit and mains are presented concurrently to determine if the 4-
higher cognitive functions. It would suggest, as various chunk limit is shared across domains or is reproduced intact
commentators have proposed, that the capacity limit may within each domain even in this dual-task situation. We
reflect the optimization of certain quantities, such as the have plenty of dual-task experiments but most have not at-
necessary degree of overlap between the neural represen- tempted to use tasks that estimate pure STM capacity. Sec-
tations of different concepts (Usher et al.). ond, we need studies of the correlations of pure STM ca-
The other basic kind of neurobiological research, includ- pacity estimates across domains.
ing electrophysiological recording and neuroimaging, is es- Lacking this evidence, I will still offer an opinion because
sentially correlative in nature rather than manipulative (if anyone who has followed this extensive controversy has
carried out only on normal adult humans to ascertain group earned some tentative bottom line. I lean toward H1, in
means). Thus, one tries to observe correlations between which the focus of attention is limited in capacity, essen-
certain patterns of behaviors and patterns of brain re- tially because we have seen that very different underlying

References/Cowan: The magical number 4
forms of memory representation yield similar capacity esti- Allport, D. A. (1968) Phenomenal simultaneity and the perceptual moment
mates. For example, consider two situations that may be hypothesis. British Journal of Psychology 59:395406. [rNC]
(1987) Selection-for-action: Some behavioural and neurophysiological
best suited to estimate capacity: those in which only one de- considerations of attention and action. In: Perspectives on perception and
cision has had to be made per trial and there has been no action, ed. H. Heuer & A. F. Sanders. Erlbaum. [TS]
interference from previous outputs (e.g., Cowan et al. un- Amit, D. J. (1995) The Hebbian paradigm reintegrated: Local reverberations as
published, discussed in sect. R2.3.1; Luck & Vogel 1997). internal representations. Behavioral and Brain Sciences 18:615 57. [AR]
The unattended-speech research shows a bow-shaped ser- Anderson, J. R. (1983) The architecture of cognition. Harvard University Press.
[SM]
ial position function, whereas running memory span shows Anderson, J. R. & Lebiere, C. (1998) The atomic components of thought. Erlbaum.
a monotonic increasing function across the tested serial po- [NAT]
sitions; and the task of Luck and Vogel is based on spatial Anderson, J. R. & Matessa, M. (1997) A production system theory of serial
arrays of colors represented simultaneously. Despite such memory. Psychological Review 104:72848. [aNC, SM, rNC]
Anderson, R. B., Tweney, R. D., Rivardo, M. & Duncan, S. (1997) Need
representational differences, a common capacity limit seems probability affects retention: A direct demonstration. Memory and Cognition
to apply, and it requires explanation. 25:86772. [PM]
The best suggestion I can offer is that aspects of the mem- Ashby, F. G., Tien, J.-Y. & Balikrishnan, J. D. (1993) Response time distributions in
ory representation determine what chunks will be most memory scanning. Journal of Mathematical Psychology 37:526 55. [rNC]
prominent (relative to the available retrieval context), Atkinson, J., Campbell, F. W. & Francis, M. R. (1976a) The magic number 4 6 0:
A new look at visual numerosity. Perception 5:32734. [aNC, BJS]
whereas limits in the focus of attention determine how many Atkinson, J., Francis, M. R. & Campbell, F. W. (1976b) The dependence of the
of the most prominent chunks in the representation can be visual numerosity limit on orientation, colour, and grouping of the stimulus.
attended at once. Thus, for example, in memory for unat- Perception 5:33542. [aNC]
tended speech, about four prominent items typically would Atkinson, R. C. & Shiffrin, R. M. (1968) Human memory: A proposed system and
its control processes. In: The psychology of learning and motivation: Advances
be picked off from the two ends of the list together; in research and theory, vol. 2, ed. K. W. Spence & J. T. Spence. Academic
whereas, in running span, about four prominent items typ- Press. [aNC, rNC]
ically would be picked off from the recency portion alone. Avons, S. E., Wright, K. L. & Pammer, K. (1994) The word-length effect in probed
We do not know which portions of a spatial array are most and serial recall. Quarterly Journal of Experimental Psychology 47A:207 31.
prominent in memory but there probably are especially [aNC]
Baars, B. J. (1988) A cognitive theory of consciousness. Cambridge University
prominent areas, perhaps those nearest the fovea. Despite Press. [aNC, BJB, rNC]
the representational diversity, a common capacity limit (1993) How does a serial, integrated and very limited stream of consciousness
seems to apply. emerge from a nervous system that is mostly unconscious, distributed, parallel
Some concepts suggested by commentaries may lead to and of enormous capacity? Theoretical and Experimental Studies of
Consciousness, Ciba Foundation Symposium 174:282303. [BJB]
minor modifications in this H1 account. Suppose Usher et (1997) In the theater of consciousness: The workspace of the mind. Oxford
al.s concept of representational overlap does not apply for University Press. [BJB]
activated representations in general, but only for a special (1998) Metaphors of consciousness and attention in the brain. Trends in
type of representation that includes just recently-attended Neurosciences 21:5862. [BJB]
items. It may be within that type of representation only that Baars, B. J. & Newman, J., ed. (in press) Essential sources in the scientific study of
consciousness. MIT Press/Bradford Books. [BJB]
representational overlap limits the capacity to 3 to 5 chunks. Bachelder, B. L. (1978) Span Theory: Laboratory and classroom applications.
If so, this could be counted either as a limit in the focus of Paper presented at the Eleventh Annual Gatlinburg Conference on Research
attention per se (H1), or as a special representational limit in Mental Retardation, Gatlinburg, TN. [BLB]
(H2). Similarly, suppose the chunk limit occurs only in a (2000) The magical number seven: What did Miller really prove? (submitted).
[BLB]
representation corresponding to Baddeleys episodic buf- Bachelder, B. L. & Denny, M. R. (1977a) A theory of intelligence: I. Span and the
fer, and not in the simpler passive stores, and suppose this complexity of stimulus control. Intelligence 1:12750. [BLB]
episodic buffer receives input only from the focus of atten- (1977b) A theory of intelligence: II The role of span in a variety of intellectual
tion. If so, it is difficult to know exactly how to apply the the- tasks. Intelligence 1:23756. [BLB]
oretical distinction shown in Figure R2. As just mentioned, Baddeley, A. D. (1986) Working memory. Clarendon Press. [aNC, ADB, SG, SM,
BR, BJS, TS, rNC]
I prefer to think that the situation best matches H1. While (1993) Visual and verbal subsystems of working memory. Current Biology
the present theoretical approach cannot yet indicate exactly 3:56365. [BJB]
the right model, it can provide a more detailed vision of the (1996) Exploring the central executive. Quarterly Journal of Experimental
limited-capacity process than before, with the benefit of an Psychology 49A:528. [ADB]
(2000) The episodic buffer: A new component for working memory? Trends in
extraordinarily thoughtful set of commentaries. Cognitive Sciences 4:417 23. [ADB]
(in press) Levels of working memory. In: Perspectives on human memory and
ACKNOWLEDGMENTS cognitive aging: Essays in honor of Fergus Craik, ed. M. Naveh-Benjamin, M.
This project was supported by NICHD Grant R01 21338. I thank Moscovitch & H. L. Roediger. Psychology Press. [rNC]
Jeffrey Rouder for helpful comments. Address correspondence to Baddeley, A. D. & Ecob, J. R. (1973) Reaction time and short-term memory:
Nelson Cowan, Department of Psychology, University of Mis- Implications of repetition effects for the high-speed exhaustive scan
souri, 210 McAlester Hall, Columbia, MO 65211. E-mail: hypothesis. Quarterly Journal of Experimental Psychology 25:229 40. [ JJ]
Baddeley, A. D., Grant, S., Wight, E. & Thompson, N. (1975a) Imagery and visual
cowanN@Missouri.edu.
working memory. In: Attention and performance V, ed. P. Rabbit and S.
Dornic. Academic Press. [ADB]
Baddeley, A. & Hitch, G. J. (1974) Working memory. In: Recent advances in
learning and motivation, vol. 8, ed. G. Bower. Academic Press. [ADB, BR]
Baddeley, A. D., Lewis, V., Eldridge, M. & Thomson, N. (1984) Attention and
References retrieval from long-term memory. Journal of Experimental Psychology:
General 113:51840. [ADB]
Letters a and r appearing before authors initials refer to target article Baddeley, A., Lewis, V. & Vallar, G. (1984) Exploring the articulatory loop. The
and response, respectively. Quarterly Journal of Experimental Psychology 36A:23352. [aNC]
Baddeley, A. D. & Logie, R. (1999) Working memory: The multiple component
Adams, D. (1985) The hitch-hikers guide to the galaxy: The original radio scripts. model. In: Models of working memory: Mechanisms of active maintenance and
Pan Books. [ JNT] executive control, ed. A. Miyake & P. Shah. Cambridge University Press. [ADB]

Baddeley, A. D., Thomson, N. & Buchanan, M. (1975b) Word length and the Cavanagh, J. P. (1972) Relation between the immediate memory span and the
structure of short-term memory. Journal of Verbal Learning and Verbal memory search rate. Psychological Review 79:52530. [DJM]
Behavior 14:575 89. [aNC, SM] Charniak, E. (1993) Statistical language learning. MIT Press. [EMP]
Basso, A., Spinnler, H., Vallar, G. & Zanobio, E. (1982) Left hemisphere damage Chater, N. & Christiansen, M. (1999) Connectionist and natural language
and selective impairment of auditory verbal STM. Neuropsychologia 20:263 processing. In: Language Processing, ed. S. Garrod & M. Pickering.
74. [ADB] Psychology Press. [EMP]
Bender, S., Schall, U., Wolstein, J., Grzella, I., Zerbin, D. & Oades, R. D. (1999) A Chase, W. G. & Ericsson, K. A. (1981) Skilled memory. In: Cognitive skills and
topographic event-related potential follow-up study of prepulse inhibition in their acquisition, ed. J. R. Anderson. Erlbaum. [KAE]
first and second episode patients with schizophrenia. Psychiatry Research (1982) Skill and working memory. In: The psychology of learning and
Neuroimaging 90:41 53. [RDO] motivation, vol. 16, ed. G. H. Bower. Academic Press. [KAE]
Besner, D. (1987) Phonology, lexical access in reading, and articulatory Chase, W & Simon, H. A. (1973) The mind eyes in chess. In: Visual information
suppression: A critical review. Quarterly Journal of Experimental Psychology processing, ed. by W. G. Chase. Academic Press. [aNC, PCL, rNC]
39A:46778. [aNC] Chelazzi, L., Miller, E. K., Duncan, J. & Desimone, R. (1993) A neural basis for
Biro, D. & Matsuzawa, T. (1999) Numerical ordering in a chimpanzee (Pan visual search in inferior temporal cortex. Nature 363:34547. [PRR]
troglodytes): Planning, executing, and monitoring. Journal of Comparative Chen, S., Swartz, K. B. & Terrace, H. S. (1998) Knowledge of the ordinal position
Psychology 113:178 85. [NK] of list items in rhesus monkeys. Psychological Science 8:8086. [DT]
Bjork, R. A. & Whitten, W. B. (1974) Recency-sensitive retrieval processes in long- Cherry, E. C. (1953) Some experiments on the recognition of speech, with one and
term free recall. Cognitive Psychology 6:17389. [aNC] with two ears. The Journal of the Acoustical Society of America 25:975 79.
Boudewijnse, G.-J. A., Murray, D. J. & Bandomir, C. A. (1999) Herbarts [aNC, HT]
mathematical psychology. History of Psychology 2:16393. [DJM] Cheung, H. & Kemper, S. (1993) Recall and articulation of English and Chinese
Bower, G. H. & Winzenz, D. (1969) Group structure, coding and memory for digit words by Chinese-English bilinguals. Memory and Cognition 21:666 70.
series. Journal of Experimental Psychology Monographs 80:117. [aNC] [SM]
Brandimonte, M. A., Hitch, G. J. & Bishop, D. V. M. (1992) Influences of short- Chi, M. T. H. & Klahr, D. (1975) Span and rate of apprehension in children and
term memory codes on visual image processing: Evidence from image adults. Journal of Experimental Child Psychology 19:43439. [aNC]
transformation tasks. Journal of Experimental Psychology: Learning, Memory, Chomsky, N. (1975) Reflections on language. Pantheon. [EMP]
and Cognition 18:157 65. [SM] Cleeremans, A. & McClelland, J. L. (1991) Learning the structure of event
Braun, H. A., Wissing, H., Schfer, K. & Hirsch, M. C. (1994) Oscillation and noise sequences. Journal of Experimental Psychology: General 120:23553. [aNC,
determine signal transduction in shark multimodal sensory cells. Nature EMP]
367:270 73. [aNC] Clifton, C & Cruse, D. (1977) Time to recognize tones: Memory scanning or
Brebion, G., Smith, M. J., Gorman, J. M., Malaspina, D., Sharif, Z. & Amador, X. memory strength? Quarterly Journal of Experimental Psychology 29:709 26.
(2000) Memory and schizophrenia: Differential link of processing speed and [rNC]
selective attention with two levels of encoding. Journal of Psychiatric Cohen, J. D., Braver, T. S. & OReilly, R. C. (1996) A computational approach to
Research 34:12127. [RDO] prefrontal cortex, cognitive control and schizophrenia: Recent developments
Bregman, A. S. & Campbell, J. (1971) Primary auditory stream segregation and and current challenges. Philosophical Transactions of the Royal Society of
perception of order in rapid sequences. Journal of Experimental Psychology London B 351:151527. [MU]
89:244 49. [rNC] Cohen, J. D., Forman, S. D., Braver, T. S., Casey, B. J., Servan-Schreiber, D. &
Bregman, A. S. & Dannenbring, G. L. (1973) The effect of continuity on auditory Noll, D. C. (1994) Activation of the prefrontal cortex in a nonspatial working
stream segregation. Perception and Psychophysics 13:308312. [rNC] memory task with function MRI. Human Brain Mapping 293304. [BR]
Broadbent, D. E. (1958) Perception and communication. Pergamon Press. [aNC, Cohen, J. D., Perlstein, W. M., Braver, T. S., Nystrom, L. E., Noll, D. C., Jonides,
HT, rNC] J. & Smith, E. E. (1977) Temporal dynamics of brain activation during a
Broadbent, D. E. (1971) Decision and stress. Academic Press. [rNC] working memory task. Nature 386:604608. [aNC]
(1975) The magic number seven after fifteen years. In: Studies in long-term Conrad, R. (1965) Order error in immediate serial recall of sequences. Journal of
memory, ed. A. Kennedy & A. Wilkes. Wiley. [aNC] Verbal Learning and Verbal Behavior 4:16169. [SEA]
Brown, G. D. A. & Hulme, C. (1995) Modeling item length effects in memory Corballis, M. C. (1967) Serial order in recognition and recall. Journal of
span: No rehearsal needed? Journal of Memory and Language 34:594621. Experimental Psychology 74:99105. [aNC]
[aNC, SM] Corbetta, M., Miezin, F. M., Dobmeyer, S., Shulman, G. L. & Petersen, S. E.
Brown, G. D. A., Preece, T. & Hulme, C. (2000) Oscillator-based memory for (1991) Selective and divided attention during visual discrimination of shape,
serial order. Psychological Review 107:12781. [aNC, rNC] color, and speed: Functional anatomy by positron emission tomography. The
Burgess, N. & Hitch, G. J. (1997) Memory for serial order: A network model of the Journal of Neuroscience 11:23832402. [BR]
phonological loop and its timing. Psychological Review 106:55181. [SM] Coslett, H. B. & Saffran, E. (1991) Simultanagnosia: To see but not two see. Brain
Burle, B. & Bonnet, M. (2000) High-speed memory scanning: A behavioral 114:152345. [BJS]
argument for a serial oscillatory model. Cognitive Brain Research 9:32737. Cowan, N. (1984) On short and long auditory stores. Psychological Bulletin
[OJ, rNC] 96:34170. [rNC]
Burrows, D. & Okada, R. (1975) Memory retrieval from long and short lists. (1988) Evolving conceptions of memory storage, selective attention, and their
Science 188:1031 33. [JJ] mutual constraints within the human information processing system.
Burtis, P. J. (1982) Capacity increase and chunking in the development of short- Psychological Bulletin 104:16391. [aNC, SM, HT, rNC]
term memory. Journal of Experimental Child Psychology 34:387413. [JPL] (1995) Attention and memory: An integrated framework. Oxford Psychology
Caplan, D., Rochon, E. & Waters, G. S. (1992) Articulatory and phonological Series, No. 26. Oxford University Press. [aNC, BJS, HT, rNC]
determinants of word length in span tasks. Quarterly Journal of Experimental (1999) An embedded-processes model of working memory. In: Models of
Psychology 45A:177 92. [SM] working memory: Mechanisms of active maintenance and executive control,
Cardozo, B. L. & Leopold, F. F. (1963) Human code transmission. Ergonomics eds. A. Miyake & P. Shah. Cambridge University Press. [rNC]
6:133 41. [aNC] Cowan, N., Cartwright, C., Winterowd, C. & Sherk, M. (1987) An adult model of
Carey, S. & Xu, F. (in press) Infant knowledge of objects: Beyond object files and preschool childrens speech memory. Memory and Cognition 15:51117.
object tracking. Cognition: Special Issue on Objects and Attention. [BJS] [aNC]
Case, R. (1974) Structures and strictures: Some functional limitations on the Cowan, N., Day, L., Saults, J. S., Keller, T. A., Johnson, T. & Flores, L. (1992) The
course of cognitive development. Cognitive Psychology 6:54473. [SM] role of verbal output time in the effect of word length on immediate memory.
(1985) Intellectual development: Birth to adulthood. Academic Press. [SM] Journal of Memory and Language 31:117. [SM]
(1995) Capacity-based explanations of working memory growth: A brief history Cowan, N., Lichty, W. & Grove, T. R. (1990) Properties of memory for unattended
and reevaluation. In: Memory performance and competencies: Issues in spoken syllables. Journal of Experimental Psychology: Learning, Memory, and
growth and development, ed. F. E. Weinert & W. Schneider. Erlbaum. [SM] Cognition 16:25869. [rNC]
(1998) The development of conceptual structures. In: Handbook of child Cowan, N., Nugent, L. D., Elliot, E. M. & Geer, T. (2000a) Is there a temporal
psychology, vol. 2, Cognition, perception, and language, ed. W. Damon, D. basis of the word length effect? A response to Service (1998). Quarterly
Kuhn & R. S. Siegler. John Wiley and Sons. [JPL] Journal of Experimental Psychology 53A:64760. [rNC]
Case, R. & Okamoto, Y. (1996) The role of central conceptual structures in the Cowan, N., Nugent, L. D., Elliot, E. M., Ponomarev, I. & Saults, J. S. (1999) The
development of childrens thought. Monographs of the Society for Research in role of attention in the development of short-term memory: Age differences in
Child Development, vol. 61, no. 245. [JPL] the verbal span of apprehension. Child Development 70:108297. [aNC,
Catchpole, C. K. & Slater, P. J. B. (1995) Bird song: Biological themes and SAH, JPL, HT, rNC]
variations. Cambridge University Press. [DT] Cowan, N., Nugent, L. D., Elliot, E. M. & Saults, J. S. (2000b) Persistence of

memory for ignored lists of digits: Areas of development constancy and Diener, D. (1990) Role of the preprobe delay in memory-scanning tasks. Memory
change. Journal of Experimental Child Psychology 76:15172. [rNC] and Cognition 18:45158. [rNC]
Cowan, N., Saults, J. S. & Nugent, L. D. (1997a) The role of absolute and relative Dirlam, D. K. (1972) Most efficient chunk sizes. Cognitive Psychology 3:355 59.
amounts of time in forgetting within immediate memory: The case of tone [aNC]
pitch comparisons. Psychonomic Bulletin and Review 4:39397. [aNC, rNC] Dosher, B. A. (1981) The effect of delay and interference: A speed-accuracy study.
(in press) The ravages of absolute and relative amounts time on memory. In: The Cognitive Psychology 13:55182. [BM, rNC]
nature of remembering: Essays in honor of Robert G. Crowder, ed. H. (1984) Degree of learning and retrieval speed: Study time and multiple
Roediger. American Psychological Association. [rNC] exposures. Journal of Experimental Psychology: Learning, Memory, and
Cowan, N., Winkler, I., Teder, W. & Ntnen, R. (1993) Memory prerequisites of Cognition 10:54174. [BM]
the mismatch negativity in the auditory event-related potential (ERP). Journal (1998) Models of visual search: Finding a face in the crowd. In: Methods,
of Experimental Psychology: Learning, Memory, and Cognition 19:90921. models, and conceptual issues: An invitation to cognitive science, ed. D.
[aNC] Scarborough and S. Sternberg. MIT Press. [rNC]
Cowan, N., Wood, N. L. & Borne, D. N. (1994) Reconfirmation of the short-term Dosher, B. A. & Ma, J. J. (1998) Output loss or rehearsal loop? Output-time versus
storage concept. Psychological Science 5:103106. [rNC] pronunciation-time limits in immediate recall for forgetting-matched
Cowan, N., Wood, N. L., Nugent, L. D. & Treisman, M. (1997b) There are two materials. Journal of Experimental Psychology: Learning, Memory, and
word length effects in verbal short-term memory: Opposed effects of duration Cognition 24:31635. [BM]
and complexity. Psychological Science 8:29095. [aNC, rNC] Duncan, J. (1984) Selective attention and the organization of visual information.
Cowan, N., Wood, N. L., Wood, P. K., Keller, T. A., Nugent, L. D. & Keller, C. V. Journal of Experimental Psychology: General 113:50117. [GD]
(1998) Two separate verbal processing rates contributing to short-term Elliot, M. A. & Mller, H. J. (1998) Synchronous information presented in 40 Hz
memory span. Journal of Experimental Psychology: General 127:14160. flicker enhances visual feature binding. Psychological Science 9:277 83.
[aNC] [rNC]
Craik, F. I. M. & Birtwistle, J. (1971) Proactive inhibition in free recall. Journal of (2000) Evidence for a 40-Hz oscillatory short-term visual memory revealed by
Experimental Psychology 91:12023. [rNC] human reaction-time measurements. Journal of Experimental Psychology:
Craik, F., Gardiner, J. M. & Watkins, M. J. (1970) Further evidence for a negative Learning, Memory, and Cognition 26:70318. [rNC]
recency effect in free recall. Journal of Verbal Learning and Verbal Behavior Elman, J. L. (1993) Learning and development in neural networks: The
9:554 60. [aNC] importance of starting small. Cognition 48:7199. [aNC, EMP]
Craik, F. I. M., Govoni, R., Naveh Benjamin, M. & Anderson, N. D. (1996) The (1996) Rethinking innateness: A connectionist perspective on development. MIT
effects of divided attention on encoding and retrieval processes in human Press. [EMP]
memory. Journal of Experimental Psychology: General 125:15980. [ADB] Engel, A. K., Fries, P., Roelfsema, P. R., Knig, P., Brecht, M. & Singer, W. (1999)
Craik, F. I. M. & Lockhart, R. S. (1972) Levels of processing: A framework for Temporal binding, binocular rivalry, and consciousness. Consciousness and
memory research. Journal of Verbal Learning and Verbal Behavior 11:671 Cognition 8: 12851. [BJB]
84. [PM] Engel, A. K., Knig, P., Kreiter, A. K., Schillen, T. B. & Singer, W. (1992) Temporal
Crick, F. (1984) Function of the thalamic reticular complex: The searchlight coding in the visual cortex: New vistas on integration in the nervous system.
hypothesis. Proceedings of the National Academy of Sciences, USA 81:4586 Trends in Neuroscience 15:21826. [AR]
90. [BJB] Engle, R. W., Kane, M. J. & Tuholski, S. W. (1999) Individual differences in
(1994) The astonishing hypothesis. Touchtone Books. [JPL] working memory capacity and what they tell us about controlled attention,
Crowder, R. G. (1989) Imagery for musical timbre. Journal of Experimental general fluid intelligence, and functions of the prefrontal cortex. In: Models of
Psychology: Human Perception and Performance 15:47278. [rNC], working memory: Mechanisms of active maintenance and executive control,
(1993) Short-term memory: Where do we stand? Memory and Cognition ed. A. Miyake and P. Shah. Cambridge University Press. [aNC]
21:142 45. [aNC] Ericsson, K. A. (1985) Memory skill. Canadian Journal of Psychology 39:188 231.
Cunningham, T. F., Healy, A. F., Till, R. E., Fendrich, D. W. & Dimitry, C. Z. [aNC, HT]
(1993) Is there really very rapid forgetting from primary memory? The role of (1996) The road to excellence: The acquisition of expert performance in the arts
expectancy and item importance in short-term recall. Memory and Cognition and sciences, sports, and games. Erlbaum. [KAE]
21:671 88. [PM] Ericsson, K. A., Chase, W. G. & Faloon, S. (1980) Acquisition of a memory skill.
Damasio, A. R. (1989) Time-locked multiregional retroactivation: A systems-level Science 208:118182. [aNC, HT, rNC]
proposal for the neural substrates of recall and recognition. Cognition 33:25 Ericsson, K. A. & Kintsch, W. (1995) Long-term working memory. Psychological
62. [BJB] Review 102:21145. [aNC, KAE]
Daneman, M. & Carpenter, P. A. (1980) Individual differences in working memory Ericsson, K. A., Patel, V. L. & Kintsch, W. (2000) How experts adaptation to
and reading. Journal of Verbal Learning and Verbal Behavior 19:45066. representative task demands account for expertise effect in memory recall:
[aNC, JNT, rNC] Comment on Vincente & Wang (1998). Psychological Review 107:578 92.
Daneman, M. & Merikle, P. M. (1996) Working memory and language [KAE]
comprehension: A meta-analysis. Psychonomic Bulletin and Review 3:42233. Ericsson, K. A. & Polson, P. G. (1988) Memory for restaurant orders. In: The
[aNC, rNC] nature of expertise, ed. M. Chi, R. Glaser & M. Farr. Erlbaum. [KAE]
Darwin, C. J., Turvey, M. T. & Crowder, R. G. (1972) An auditory analogue of the Fabiani, M., Stadler, M. A. & Wessels, P. M. (in press) True memories but not false
Sperling partial report procedure: Evidence for brief auditory storage. ones produce a sensory signature in human lateralized brain potentials.
Cognitive Psychology 3:25567. [aNC, rNC] Journal of Cognitive Neuroscience. [GG]
Davis, G., Driver, J., Pavani, F. & Shepherd A. J. (2000) Reappraising the costs of Feigenbaum, E. A. & Simon, H. A. (1984) EPAM-like models of recognition and
attending to two separate visual objects. Vision Research 40:132332. [GD] learning. Cognitive Science. 8:30536. [PCL]
Davis, G., Welch, V. L., Holmes, A. & Shepherd, A. J. (submitted) Can attention Fisher, D. L. (1984) Central capacity limits in consistent mapping, visual search
select only a fixed number of objects at a time? tasks: Four channels or more? Cognitive Psychology 16:449 84. [aNC, TS]
DEsposito, M., Detre, J., Alsop, D., Shin, R., Atlas, S. & Grossman, M. (1995) The Fraise, P. (1978) Time and rhythm perception. In: Handbook of Perception, vol. 8,
neural basis of the central executive system of working memory. Nature eds. E. Carterette & M. Friedman. Academic Press. [SG]
378:279 81. [BR] Frensch, P. A. & Miner, C. S. (1994) Effects of presentation rate and individual
De Groot, A. D. (1978) Thought and choice in chess, 2nd ed. Morton. [PCL] differences in short-term memory capacity on an indirect measure of serial
Dehaene, S. & Cohen, L. (1994) Dissociable mechanisms of subitizing and learning. Memory and Cognition 22:95110. [aNC]
counting: Neuropsychological evidence from simultanagnosic patients. Frick, R. W. (1984) Using both an auditory and a visual short-term store to increase
Journal of Experimental Psychology: Human Perception and Performance digit span. Memory and Cognition 12:50714. [aNC]
20:958 75. [BJS] Friedman-Hill, S. R., Robertson, L. C. & Treisman, A. (1995) Parietal
Dempster, F. N. (1981) Memory span: Sources of individual and developmental contributions to visual feature binding: Evidence from a patient with bilateral
differences. Psychological Bulletin 89:63100. [aNC] lesions. Science 269:85355. [BJS]
Desimone, R. (1998) Visual attention mediated by biased competition in Frith, C. D. & Dolan, R. J. (1997) Brain mechanisms associated with top-down
extrastriate visual cortex. Philosophical Transactions of the Royal Society of processes in perception. Philosophical Transactions of the Royal Society of
London B 353:124555. [MU] London B, 352:122130. [RDO]
Desimone, R. & Duncan, J. (1995) Neural mechanisms of selective visual Fuster, J. M. (1997) Network memory. Trends in Neuroscience 20:451 59.
attention. Annual Review of Neuroscience 18:193222. [PRR] [PRR]
Destexhe, A., Contreras, D. & Steriade, M. (1999) Spatiotemporal analysis of local Fuster, J. M., Bauer, R. H. & Jervey, J. P. (1985) Functional interactions between
field potentials and unit discharges in cat cerebral cortex during natural wake inferotemporal and prefrontal cortex in a cognitive task. Brain Research
and sleep states. Journal of Neuroscience 19:45954608. [BJB] 330:299307. [PMM]

Gathercole, S. E. & Baddeley, A. D. (1993) Working memory and language. Halford, G. S., Wilson, W. H. & Phillips, S. (1998) Processing capacity defined by
Erlbaum. [ADB] relational complexity: Implications for comparative, developmental, and
Geary, D. C. (1993) Mathematical disabilities: Cognitive, neuropsychological, and cognitive psychology. Behavioral and Brain Sciences 21:723802. [aNC,
genetic components. Psychological Bulletin 114:34562. [SAH, rNC] GSH, JPL, rNC]
(1995) Reflections of evolution and culture in childrens cognition. American Hamilton, W. (1859) Lectures on metaphysics and logic, vol. 1. Blackwood.
Psychologist 50:24 36. [SAH, rNC] [aNC]
George, N., Jemel, B., Fiori, N. & Renault, B. (2000) Holistic and part-based face Hatano, G. & Osawa, K. (1983) Digit memory of grand experts in abacus-derived
representations: Evidence from the memory span of the face superiority mental calculation. Cognition 15:95110. [KAE]
effect. Current Psychology Letters 1:89106. [RDO] Hausman, A. & Wilson, F. (1967). Goodmans ontology. In: Carnap and Goodman:
Glanzer, M. & Cunitz, A. R. (1966) Two storage mechanisms in free recall. Journal Two formalists, ed. A. Hausman & F. Wilson, Iowa Publications in Philosophy,
of Verbal Learning and Verbal Behavior 5:35160. [aNC] vol. 3. Martinus Nijhoff. [JPL]
Glanzer, M. & Razel, M. (1974) The size of the unit in short-term storage. Journal Hazeltine, E., Ivry, R. & Teague, D. (2000) Dual-task performance with minimal
of Verbal Learning and Verbal Behavior 13:11431. [aNC] cost: Evidence for concurrent response selection processes. Journal of
Glenberg, A. M. & Swanson, N. C. (1986) A temporal distinctiveness theory of Cognitive Neuroscience supplement:110. [TS]
recency and modality effects. Journal of Experimental Psychology: Learning, He, S., Cavanaugh, P. & Intriligator, J. (1997) Many objects can be enumerated in
Memory, and Cognition 12:315. [aNC] afterimages when they are not crowded [Abstract]. Abstracts of the
Gobet, F. (1998) Memory for the meaningless: How chunks help. Proceedings of Psychonomic Society 2:3. [BJS, rNC]
the Twentieth Meeting of the Cognitive Science Society. Erlbaum. [PCL] Healy, A. F. & Cunningham, T. F. (1995) Very rapid forgetting: Reply to Muter.
Gobet, F. & Simon, H. A. (1996) Templates in chess memory: A mechanism for Memory and Cognition 23:38792. [PM]
recalling several boards. Cognitive Psychology 31:140. [aNC, KAE] Hebb, D. O. (1949) The organization of behavior. Wiley. [MU]
(1998) Expert chess memory: Revisiting the chunking hypothesis. Memory Hecht, S. A. (1998) Toward an information processing account of individual
6:225 55. [aNC, PCL] differences in fraction skills. Journal of Educational Psychology 90:545 559.
(in press) Five seconds or sixty? Presentation time in expert memory. Cognitive [SAH]
Science. [PCL] Hecht, S. A., Torgesen, J. K., Wagner, R. K. & Rashotte, C. A. (in press) The
Gold, E. M. (1967) Language identification in the limit. Information and Control relations between phonological processing abilities and emerging individual
16:44774. [EMP] differences in mathematical computation skills: A longitudinal study from
Gomes, H., Sussman, E., Ritter, W., Kurtzberg, D., Cowan, N. & Vaughan Jr., H. second- to fifth-grade. Journal of Experimental Child Psychology. [SAH]
G. (1999) Electrophysiological evidence of developmental changes in the Hellige, J. B. & Cox, P. J. (1976) Effects of concurrent verbal memory on
duration of auditory sensory memory. Developmental Psychology 35:294302. recognition of stimuli from the left and right visual fields. Journal of
[rNC] Experimental Psychology: Human Perception and Performance 2:210 221.
Goodman, N. (1951) The structure of appearance. Harvard University Press. [SM]
[ JPL] Henderson, L. (1972) Spatial and verbal codes and the capacity of STM. Quarterly
Graesser II, A. & Mandler, G. (1978) Limited processing capacity constrains the Journal of Experimental Psychology 24:48595. [aNC]
storage of unrelated sets of words and retrieval from natural categories. Henry, L. (1991) The effects of word length and phonemic similarity in young
Journal of Experimental Psychology: Human Learning and Memory 4:86100. childrens short-term memory. Quarterly Journal of Experimental Psychology
[aNC] 43A:3552. [SM]
Granholm, E., Perry, W., Filoteo, J. V. & Braff, D. (1999) Hemispheric and Henry, L. (1994) The relationship between speech rate and memory span in
attentional contributions to perceptual organization deficits on the global-local children. International Journal of Behavioral Development 17:3756. [SM]
task in schizophrenia. Neuropsychology 13:27181. [RDO] Henson, R. N. A. (1999) Positional information in short-term memory: Relative or
Gratton, G., Corballis, P. M. & Jain, S. (1997) Hemispheric organization of visual absolute? Memory and Cognition 27:91527. [rNC]
memories. Journal of Cognitive Neuroscience 9:92104. [GG] Herbart, J. F. (1890) Psychologie als Wissenschaft [Psychology as science]. In: Jon.
Gratton, G., Fabiani, M., Goodman-Wood, M. R. & DeSoto, M. C. (1998) Fr. Herbarts saemtliche Werke in chronologischer Reihenfolge [Jon. Fr.
Memory-driven processing in human medial occipital cortex: An event-related Herbarts collected works in their chronological order of publication], part 1,
optical signal (EROS) study. Psychophysiology 38:34851. [GG] vol. 5, ed. K. Kehrbach & O. Flgel. Herman Beyer und Soehne. [DJM]
Gray, C. M., Knig, P., Engel, A. K. & Singer, W. (1989) Oscillatory responses in Hitch, G. J., Burgess, N., Towse, J. N. & Culpin, V. (1996) Temporal grouping
cat visual cortex exhibit inter-columnar synchronization, which reflects global effects in immediate recall: A working memory analysis. Quarterly Journal of
stimulus properties. Nature 338:33437. [aNC, PRR, HT] Experimental Psychology 49A:11639. [aNC]
Greene, R. L. (1989) Immediate serial recall of mixed-modality lists. Journal of Hobson, J. A. (1997) Consciousness as a state-dependent phenomenon. In:
Experimental Psychology: Learning, Memory, and Cognition 15:26674. Scientific approaches to consciousness: The Twenty-eighth Carnegie
[aNC] Symposium on Cognition, ed. J. D. Cohen & J. W. Schooler. Erlbaum. [BJB]
Greenwald, A. & Shulman, H. (1973) On doing two at once: II. Elimination of the Holtzman, J. D. & Gazzaniga, M. S. (1982) Dual task interactions due exclusively
psychological refractory period. Journal of Experimental Psychology 101:70 to limits in processing resources. Science 218:132527. [aNC]
76. [TS] Holyoak, K. J. & Thagard, P. (1995) Mental leaps: Analogy in creative thought.
Grondin, S. (in press) From physical time to the first and second moments of Bradford. [GSH]
psychological time. Psychological Bulletin. [SG] Horn D. & Opher I. (1996) Temporal segmentation in a neural dynamic system.
Grondin, S., Meilleur-Wells, G. & Lachance, R. (1999) When to start explicit Neural Computation 8:37388. [MU]
counting in time-intervals discrimination tasks: A critical point in the timing Horn, D. & Usher, M. (1991) Parallel activation of memories in an oscillatory
process of humans. Journal of Experimental Psychology: Human Perception neural network. Neural Computation 3:3143. [MU]
and Performance 25:9931004. [SG] (1992) An oscillatory model for short term memory. In: Advances in neural
Gross, C. G., Bender, D. B. & Rocha-Miranda, C. E. (1974) Inferotemporal cortex: information processes systems 4, ed. J. E. Moody, S. J. Hanson & P. R.
A single-unit analysis. In: The neurosciences: Third study program, ed. F. O. Lippmann. Morgan and Kaufmann. [MU]
Schmitt & F. G. Worden. MIT Press. [PMM] Horowitz, T. S. & Wolfe, J. M. (1998) Visual search has no memory. Nature
Grossberg, S. (1976) Adaptive pattern classification and universal recoding. 394:57577. [GFW]
Biological Cybernetics 23:12134. [MU] Houghton, G. & Tipper, S. P. (1994) A model of inhibitory mechanisms in selective
Gruenewald, P. J. & Lockhead, G. R. (1980) The free recall of category examples. attention. In: Inhibitory mechanisms in attention, memory, and language, ed.
Journal of Experimental Psychology: Human Learning and Memory 6:225 D. Dagenbach & T. Carr. Academic Press. [SM]
40. [aNC] Hulme, C., Maugham, S. & Brown, G. D. (1991) Memory for familiar and
Guilford, J. P. & Dallenback, K. M. (1925) The determination of memory span by unfamiliar words: Evidence for a long-term memory contribution to short
the method of constant stimuli. American Journal of Psychology 36:62128. term memory span. Journal of Memory and Language 30:685701. [SM]
[BLB] Hulme, C & Tordoff, V. (1989) Working memory development: The effects of
Guttentag, R. E. (1984) The metal effort requirement of cumulative rehearsal: A speech rate, word length and acoustic similarity on serial recall. Journal of
developmental study. Journal of Experimental Child Psychology 37:92106. Experimental Child Psychology 47:7287. [SM]
[aNC, rNC] Hultsch, H. (1991) Early experience can modify singing styles: Evidence from
Halford, G. S. (1993) Childrens understanding: The development of mental experiments with nightingales, Luscinia megarhynchos. Animal Behavior
models. Erlbaum. [SM, JPL] 42:88389. [DT]
Halford, G. S., Mayberry, M. T. & Bain, J. D. (1988) Set-size effects in primary (1992) Time window and unit capacity: Dual constraints on the acquisition of
memory: An age-related capacity limitation? Memory and Cognition 16:480 serial information in songbirds. Journal of Comparative Physiology A
87. [aNC] 170:27580. [DT]

(1993) Tracing the memory mechanisms in the song acquisition of birds. Kareev, Y., Lieberman, I. & Lev, M. (1997) Through a narrow window: Sample size
Netherlands Journal of Zoology 43:15571. [DT] and the perception of correlation. Journal of Experimental Psychology:
Hultsch, H., Mundry, R. & Todt, D. (1998) Learning, representation and retrieval General 126:27887. [aNC]
of rule-related knowledge in the song system of birds. In: Learning: Rule Kaufman, E. L., Lord, M. W., Reese, T. W. & Volkmann, J. (1949) The
extraction and representation, ed. A. Friederici & R. Menzel. Walter de discrimination of visual number. American Journal of Psychology 62:498 525.
Gruyter. [DT] [aNC, JJ]
Hultsch, H. Schleuss, F. & Todt, D. (1999) Auditory-visual stimulus pairing Kawai, N. (in press) Ordering and planning in sequential responding to arabic
enhances perceptual learning in a songbird. Animal Behavior 58:114349. numbers by a chimpanzee. Psychologia. [NK]
[DT] Kawai, N. & Matsuzawa, T. (2000a) A conventional approach to chimpanzee
Hultsch, H. & Todt, D. (1989) Memorization and reproduction of songs in cognition: Response to M. D. Hauser. Trends in Cognitive Science 4:128 29.
nightingales (Luscinia megarhynchos): Evidence for package formation. [NK]
Journal of Comparative Physiology A 165:197203. [DT] (2000b) Numerical memory span in a chimpanzee. Nature 403:39 40. [NK,
Hummel, J. E. & Holyoak, K. J. (1997) Distributed representations of structure: A rNC]
theory of analogical access and mapping. Psychological Review 104:42766. (2001) Reproductive memory processes in chimpanzees: Homologous
[aNC] approaches to research on human working memory. In: Primate origins of
Humphreys, G. W. (1998) Neural representation of objects in space: A dual coding human cognition and behavior, ed. T. Matsuzawa. Springer-Verlag. [NK]
account. Philosophical Transactions of the Royal Society of London B Keller, T. A. & Cowan, N. (1994) Developmental increase in the duration of
353:1341 51. [GD] memory for tone pitch. Developmental Psychology 30:855 63. [rNC]
Irwin, D. E. (1992) Memory for positions and identity across eye movements. Keller, T. A., Cowan, N. & Saults, J. S. (1995) Can auditory memory for tone pitch
Journal of Experimental Psychology: Learning, Memory, and Cognition be rehearsed? Journal of Experimental Psychology: Learning, Memory, and
18:30717. [WS] Cognition 21:63545. [rNC]
Irwin, D. E. & Yeomans, J. M. (1986) Sensory registration and informational Killeen, P. R. & Weiss, N. A. (1987) Optimal timing and the Weber function.
persistence. Journal of Experimental Psychology: Human Perception and Psychological Review 94:45568. [SG]
Performance 12:34360. [rNC] Kintsch, W. & van Dijk, T. A. (1978) Toward a model of text comprehension and
Jacoby, L. L. & Bartz, W. H. (1972) Encoding processes and the negative recency production. Psychological Review 85:36394. [aNC]
effect. Journal of Verbal Learning and Verbal Behavior 11:56165. [PM] Kirschfeld, K. (1992) Oscillations in the insect brain: Do they correspond to
Jacoby, L. L., Woloshyn, V. & Kelley, C. (1989) Becoming famous without being theoretical (g-waves of vertebrates? Proceedings of the National Academy of
recognized: Unconscious influences of memory produced by divided Sciences 89:476468. [aNC]
attention. Journal of Experimental Psychology: General 118:11525. [aNC] Klahr, D. (1973) Quantification processes. In: Visual information processing, ed.
Jahnke, J. C., Davis, S. T. & Bower, R. E. (1989) Position and order: Information in W. G. Chase. Academic Press. [JJ]
recognition memory. Journal of Experimental Psychology: Learning, Memory Klapp, S. T. & Netick, A. (1988) Multiple resources for processing and storage in
and Cognition 15:85967. [SEA] short-term working memory. Human Factors 30:61732. [aNC]
James, W. (1890) The principles of psychology. Henry Holt. [aNC] Kleinberg, J. & Kaufman, H. (1971) Constancy for short-term memory: Bits and
Jarvella, R. J. (1971) Syntactic processing of connected speech. Journal of Verbal chunks. Journal of Experimental Psychology 90:32633. [rNC]
Language and Verbal Behavior 10:40916. [EMP] Kobatake, E. & Tanaka, K. (1994) Neuronal selectivities to complex object features
Javitt, D. C., Strous, R., Grochowski, S., Ritter, W. & Cowan, N. (1997) Impaired in the ventral visual pathway of the macaque cerebral cortex. Journal of
precision, but normal retention, of auditory sensory (Aechoic@) memory Neurophysiology 71:85667. [PRR}
information in schizophrenia. Journal of Abnormal Psychology 106:31524. Koehler, O. (1954) Vorbedingungen und Vorstufen unserer Sprache bei Tieren.
[rNC] Zoologischer Anzeiger Supplement 18:32741. [DT]
Jensen, E. M., Reese, E. P. & Reese, T. W. (1950) The subitizing and counting of Kolb, B. & Whishaw, I. Q. (1996) Fundamentals of human neuropsychology. W. H.
visually presented fields of dots. Journal of Psychology 30:36392. [JJ] Freeman and Company. [JPL]
Jensen, O. & Lisman, J. E. (1996) Novel lists of 7 6 2 known items can be reliably Konishi, M. (1989) Bird song for neurobiologists. Neuron 3:541 49. [DT]
stored in an oscillatory short-term memory network: Interaction with long- Kroodsma, D. E. & Miller, E. H. (1996) Ecology and evolution of acoustic
term memory. Learning and Memory 3:25763. [OJ] communication in birds. Cornell University Press. [DT]
(1998) An oscillatory short-term memory buffer model can account for data on Laming, D. (1992) Analysis of short-term retention: Models for Brown-Peterson
the Sternberg task. Journal of Neuroscience 18:10,68810,699. [OJ] experiments. Journal of Experimental Psychology: Learning, Memory, and
Jensen, O. & Tesche, C. D. (2000) Frontal midline activity in the theta band (6 8 Cognition 18:134265. [PM]
Hz) increases with memory load in a short-term memory task: A parametric Lane, P. C. R., Gobet, F. & Cheng, P. C.-H. (2000a) Learning perceptual chunks in
MEG study. European Research Conference, May 2000, Grenada. [OJ] a computational model of problem solving with diagrams. In: Proceedings of
Jevons, W. S. (1871) The power of numerical discrimination. Nature 3:28182. the Third International Conference on Cognitive Modeling. Universal Press.
[aNC] [PCL]
Jiang, Y., Olson, I. R. & Chun, M. M. (2000) Organization of visual short-term (2000b) Learning-based constraints on schemata. In: Proceedings of the
memory. Journal of Experimental Psychology: Learning, Memory, and Twentieth Annual Meeting of the Cognitive Science Society. Erlbaum.
Cognition 26:683702. [KAE] [PCL]
Jinks, A. & Laing, D. G. (1999) Temporal processing reveals a mechanism for LaPointe, L. B. & Engle, R. W. (1990) Simple and complex word spans as
limiting the capacity of humans to analyze odor mixtures. Cognitive Brain measures of working memory capacity. Journal of Experimental Psychology:
Research 8:31125. [rNC] Learning, Memory, and Cognition 16:111833. [aNC]
Johnson, J., Fabian, V. & Pascual-Leone, J. (1989) Quantitative hardware-stages Lavie, N. & Driver, J. (1996) On the spatial extent of attention in object-based
that constrain language development. Human Development 32:24571. visual selection. Perception and Psychophysics 58:123851. [GD]
[ JPL] Lee, C. L. & Estes, W. K. (1981) Item and order information in short-term
Johnston, J. C., McCann, R. S., & Remington, R. W. (1995) Chronometric memory: Evidence for multilevel perturbation processes. Journal of
evidence for two types of attention. Psychological Science 6:36569. [GFW] Experimental Psychology: Human Learning and Memory 7:149 69. [rNC]
Jones, D., Farrand, P., Stuart, G. & Morris, N. (1995) Functional equivalence of Leventhal, A. G., Thompson, K. F., Liu, D., Zhou, Y. & Ault, S. J. (1995)
verbal and spatial information in serial short-term memory. Journal of Concomitant sensitivity to orientation, direction, and color of cells in layers 2,
Experimental Psychology: Learning, Memory, and Cognition 21:100818. 3, and 4 of monkey striate cortex. Journal of Neuroscience 15:1808 18.
[aNC] [PRR]
Jou, J. (1998) Two retrieving modes in memory of the U.S. state names. Levy, B. A. (1971) Role of articulation in auditory and visual short-term memory.
Unpublished data. [JJ] Journal of Verbal Learning and Verbal Behavior 10:123132.
Jou, J. & Aldridge, J. W. (1999) Memory representation of alphabetic position and Lewicki, P., Czyzewska, M. & Hoffman, H. (1987) Unconscious acquisition of
interval information. Journal of Experimental Psychology: Learning, Memory, complex procedural knowledge. Journal of Experimental Psychology 13:523
and Cognition 25:680701. [JJ] 30. [aNC]
Just, M. A., Carpenter, P. A. & Hemphill, D. D. (1996) Constraints on processing Libet, B., Alberts, W. W., Wright, E. W., Delattre, L., Levin, G. & Feinstein, B.
capacity: Architectural or implementational? In: Mind matters: A tribute to (1964) Production of threshold levels of conscious sensation by electrical
Allen Newell, ed. D. M. Steier & T. M. Mitchell. Erlbaum. [CPB] stimulation of human somatosensory cortex. Journal of Neurophysiology
Kareev, Y. (1995) Through a narrow window: Working memory capacity and the 27:54678. [RDO]
detection of covariation. Cognition 56:26369. [aNC, SAH] Lisman, J. E. & Idiart, M. A. P. (1995) Storage of 7 6 2 short-term memories in
(2000) Seven (indeed, plus or minus two) and the detection of correlations. oscillatory subcycles. Science 267:151215. [aNC, OJ, PMM, AR, PRR, HT,
Psychological Review 107:397402. [rNC] MU, rNC]

Livingstone, M. S. (1996) Oscillatory firing and interneuronal correlations in (1993) Serial retrieval processes in the recovery of order information. Journal of
squirrel monkey striate cortex. Journal of Neurophysiology 75:246785. PRR] Experimental Psychology: General 122:291315. [BM]
Logan, G. D. (1988) Toward an instance theory of automatization. Psychological McGeoch, J. A. (1932) Forgetting and the law of disuse. Psychological Review
Review 95:492 527. [aNC, SAH] 39:35270. [aNC]
Logan, G. D. & Klapp, S. T. (1991) Automatizing alphabet arithmetic: I. Is McKone, E. (1995) Short-term implicit memory for words and nonwords. Journal
extended practice necessary to produce automaticity? Journal of Experimental of Experimental Psychology: Learning, Memory and Cognition 21:1108 26.
Psychology: Learning, Memory, and Cognition 17:17995. [aNC] [aNC, SEA, EM]
Logie, R. H. & Baddeley, A. D. (1987) Cognitive processes in counting. Journal of McKone, E. & Dennis, C. (2000) Short-term implicit memory: Visual, auditory,
Experimental Psychology: Learning, Memory, and Cognition 13:31026. and cross-modality priming. Psychonomic Bulletin and Review 7:341 46.
[aNC, JJ] [EM]
Logie, R. H., Gilhooly, K. J. & Wynn, V. (1994) Counting on working memory in McLean, R. S. & Gregg, L. W. (1967) Effects of induced chunking on temporal
arithmetic problem solving. Memory and Cognition 22:395410. [aNC] aspects of serial recitation. Journal of Experimental Psychology 74:455 59.
Longoni, A. M., Richardson, J. T. E. & Aiello, A. (1993) Articulatory rehearsal and [aNC]
phonological storage in working memory. Memory and Cognition 21:1122. Melton, A. W. (1963) Implications of short-term memory for a general theory of
[aNC] memory. Journal of Verbal Learning and Verbal Behavior 2:121. [aNC]
Lovett, M. C., Reder, L. M. & Lebiere, C. (1997) Modeling individual differences Mewhort, D. J. K., Campbell, A. J., Marchetti, F. M. & Campbell, J. I. D. (1981)
in a digit working memory task. In: Proceedings of the Nineteenth Conference Identification, localization, and iconic memory: An evaluation of the bar-
of the Cognitive Science Society. Erlbaum. [NAT] probe task. Memory and Cognition 9:5067. [rNC]
Luck, S. J. & Vogel, E. K. (1997) The capacity of visual working memory for Meyer, D. E. & Kieras, D. E. (1997) A computational theory of executive processes
features and conjunctions. Nature 390:27981. [aNC, SAH, AR, PRR, WS, and multiple-task performance: Part 1. Basic mechanisms. Psychological
rNC] Review 104:365. [aNC, TS, rNC]
(1998) Response to visual and auditory working memory capacity by N. Cowan. Michon, J. (1978) The making of the present: A tutorial review. In: Attention and
Trends in Cognitive Sciences 2:7880. [aNC] Performance VII, ed. J. Requin. Erlbaum. [SG]
MacGregor, J. N. (1987) Short-term memory capacity: Limitation or optimization? Miller, G. A. (1956) The magical number seven, plus or minus two: Some limits on
Psychological Review 94:107108. [aNC] our capacity for processing information. Psychological Review 63:81 97.
Mahoney, J. V. & Ullman, S. (1988) Image chunking defining spatial building [aNC, SEA, BLB, CPB, GG, SG, JJ, BJS, HT, NAT, rNC]
blocks for scene analysis. In: Computational processes in human vision: An Milner, P. M. (1974) A model for visual shape recognition. Psychological Review
interdisciplinary perspective, ed. Z. Pylyshyn. Intellect. [PRR] 81:52135. [aNC]
Mandler, G. (1967) Organization and memory. In: The psychology of learning and (1999) The autonomous brain. Erlbaum. [PMM]
motivation, vol. 1, ed. K. W. Spence & J. T. Spence. Academic Press. [aNC] Miltner, W. H. R., Braun, C., Arnold, M., Witte, H. & Taub, E. (1999) Coherence
(1975) Memory storage and retrieval: Some limits on the reach of attention and of gamma-band EEG activity as a basis for associative learning. Nature
consciousness. In: Attention and performance V, ed. P. M. A. Rabbit & S. 397:43436. [aNC]
Dornic. Academic Press. [aNC] Miyake, A. & Shah, P. (1999) Models of working memory: Mechanisms of active
(1985) Cognitive psychology: An essay in cognitive science. Erlbaum. [aNC] maintenance and executive control. Cambridge University Press. [ADB]
Mandler, G. & Shebo, B. J. (1982) Subitizing: An analysis of its component Miyashita, Y. (1993) Inferior temporal cortex: Where visual perception meets
processes. Journal of Experimental Psychology: General 111:122. [aNC, memory. Annual Review of Neuroscience 16:24563. [PMM]
BLB, JJ, rNC] Monsell, S. (1978) Recency, immediate recognition memory, and reaction time.
March, L., Cienfuegos, A., Goldbloom, L., Ritter, W., Cowan, N. & Javitt, D. C. Cognitive Psychology 10:465501. [rNC]
(1999) Normal time course of auditory recognition in schizophrenia, despite Moran, J. & Desimone, R. (1985) Selective attention gates visual processing in the
impaired precision of the auditory sensory (Aechoic@) memory code. Journal extrastriate cortex. Science 229:78284. [PMM]
of Abnormal Psychology 108:6975. [rNC] Morra, S. (1989) Untitled presentation given at the Workshop on Working
Marler, P. (1976) Sensory templates in species-specific behavior. In: Simpler Memory, Schonloo, The Netherlands, April 2124. [SM]
networks and behavior, ed J. C. Fentress. Sinauer Associates. [DT] (1990) Why another model of working memory? Paper presented at the Fourth
(1991) Differences in behavioural development in closely related species: Bird Conference of the European Society for Cognitive Psychology, Como, Italy,
song. In: The development and integration of behaviour, ed P. Bateson. September 1519. [SM]
Cambridge University Press. [DT] (1994) Issues in working memory measurement: Testing for M capacity.
Marr, D. (1982) Vision. Freeman. [rNC] International Journal of Behavioral Development 17:14359. [SM, JPL]
Marsh, R. L., Sebrechts, M. M., Hicks, J. L. & Landau, J. D. (1997) Processing (2000) A new model of verbal short-term memory. Journal of Experimental
strategies and secondary memory in very rapid forgetting. Memory and Child Psychology 75:191227. [SM, JPL]
Cognition 25:173 81. [PM] Morra, S., Mazzoni, G. & Sava, D. (1993) Esiste un loop articolatorio di capacit
Martin, M. (1978) Memory span as a measure of individual differences in memory limitata? [Is there a time-limited articulatory loop?] Proceedings of the SIPs
capacity. Memory and Cognition 6:19498. [SEA] Research Division Conference, Rome, September 29-October 2, 94 95.
(1980) Attention to words in different modalities: Four-channel presentation [SM]
with physical and semantic selection. Acta Psychologica 44:99115. [aNC] Morra, S., Moizo, C. & Scopesi, A. (1988) Working memory (of the M. operator)
Massaro, D. W. (1972) Preperceptual images, processing time, and perceptual and the planning of childrens drawings. Journal of Experimental Child
units in auditory perception. Psychological Review 79:12445. [rNC] Psychology 46:4173. [SM]
Matsuzawa, T. (1985) Use of numbers by a chimpanzee. Nature 315:5759. [NK] Morra, S., Pascual-Leone, J., Johnson, J. & Baillargeon, R. (1991) Understanding
Maunsell, J. H. R. (1995) The brains visual world: Representation of visual targets spatial descriptions: Experimental test of a mental capacity model. In: Mental
in cerebral cortex. Science 270:76469. [PRR] images in human cognition, ed. R. H. Logie & M. Denis. North Holland.
May, P., Tiitinen, H., Ilmoniemi, R. J., Nyman, G., Taylor, J. G. & Ntnen, R. [SM]
(1999) Frequency change detection in human auditory cortex. Journal of Morra, S. & Stoffel, C. (1991) A new model of verbal short term memory and its
Computational Neuroscience 6:97118. [HT] development. Paper presented at the International Conference on Memory,
McCann, R. & Johnston, J. C. (1992) Locus of the single-channel bottleneck in Lancaster, U. K., July 1519. [SM]
dual-task performance. Journal of Experimental Psychology: Human Morra, S., Vigliocco, G. & Penello, B. (in press) M capacity as a lifespan construct:
Perception and Performance 18:47164. [TS] A study of its decrease in aging subjects. International Journal of Behavioral
McElree, B. (1996) Accessing short-term memory with semantic and phonological Development. [SM]
information: A time-course analysis. Memory and Cognition 24:17387. Moscovitch, M. & Umilt, C. (1990) Modularity and neuropsychology:
[BM] Implications for the organization of attention and memory in normal and
(1997) Working memory and the focus of attention. Paper presented at the brain-damaged people. In: Modular processes in dementia, ed. M. Schwartz.
Psychonomic Society Conference, Philadelphia, Pa, November. [BM] MIT/Bradford. [SM]
(1998) Attended and nonattended states in working memory: Accessing Murdoch Jr., B. B. & Hockley, W. E. (1989) Short-term memory for associations.
categorized structures. Journal of Memory and Language 38:22552. [BM, In: The psychology of learning and motivation, vol. 24, ed. G. H. Bower.
rNC] Academic Press. [PM]
(in press) Working memory and focal attention. Journal of Experimental Murray, D. J. (1968) Articulation and acoustic confusability in short-term memory.
Psychology: Learning, Memory, and Cognition. [BM] Journal of Experimental Psychology 78:67984. [aNC]
McElree, B. & Dosher, B. A. (1989) Serial position and set size in short-term (1995) Gestalt psychology and the cognitive revolution. Harvester Wheatsheaf.
memory: Time course of recognition. Journal of Experimental Psychology: [DJM]
General 118:346 73. [BM] Murray, D. J., Boudreau, N., Burggraf, K. K., Dobell, L., Guger, S. L., Leask, A.,

Stanford, L., Tate, T. L. & Wheeler, M. (1999) A grouping interpretation of (in press) Reflections on working memory: Are the two models complementary?
the modality effect in immediate probed recognition. Memory and Cognition Journal of Experimental Child Psychology. [JPL]
27:234 45. [DJM] Pascual-Leone, J. & Baillargeon, R. (1994) Developmental measurement of mental
Murray, D. J., Burhop, J., Centa, S., Chande, N., Oinonen, K., Thomas, T., Wilkie, attention. International Journal of Behavioral Development 17:161200.
T. & Farahmand, B. (1998) A partial matching theory of the mirror effect in [SM, JPL]
immediate probed recognition. Memory and Cognition 26:11961213. [DJM] Pascual-Leone, J. & Johnson, J. (1991) The psychological unit and its role in task
Muter, P. (1980) Very rapid forgetting. Memory and Cognition 8:17479. [PM] analysis. In: Criteria for competence: Controversies in the assessment of
(1995) Very rapid forgetting: Reply to Cunningham, Healy, Till, Fendrich and childrens ablities, ed. M. Chandler & M. Chapman. Erlbaum. [SM]
Dimitry. Memory and Cognition 23:38386. [PM] Pascual-Leone, J., Johnson, J. & Morris, S. (in preparation) A quantitative model of
Nairne, J. S. (1990) A feature model of immediate memory. Memory and mental attention in a mental span task. [JPL]
Cognition 18:25169. [rNC] Pashler, H. (1988) Familiarity and visual change detection. Perception and
(1991) Positional uncertainty in long-term memory. Memory and Cognition Psychophysics 44:36978. [aNC, rNC]
19:332 40. [aNC, JJ, rNC] (1991) Shifting visual attention and selecting motor responses: Distinct
(1992) The loss of positional certainty in long-term memory. Psychological attentional mechanism. Journal of Experimental Psychology: Human
Science 3:199 202. [aNC, JJ] Perception and Performance 17:102340. [GFW]
Nairne, J. S., Neath, I. & Serra, M. (1997) Proactive interference plays a role in the (1994) Dual-task interference in simple tasks: Data and theory. Psychological
word-length effect. Psychonomic Bulletin and Review 4:54145. [SM] Bulletin 16:22044. [TS, GFW]
Neath, I. (1998) Human memory: An introduction to research, data, and theory. Penney, C. G. (1980) Order of report in bisensory verbal short-term memory.
Brooks/Cole. [aNC] Canadian Journal of Psychology 34:19095. [aNC]
Neath, I. & Nairne, J. S. (1995) Word-length effects in immediate memory: Pepperberg, I. M. (1993) A review of the effects of social interaction on vocal
Overwriting trace decay. Psychonomic Bulletin and Review 2:42941. [aNC, learning in African grey parrots (Psittacus erithacus). Netherlands Journal of
DJM, rNC] Zoology 43:10424. [DT]
Neely, J. H. (1991) Semantic priming effects in visual word recognition: A selective Peterson, L. R. & Johnson, S. T. (1971) Some effects of minimizing articulation on
review of current findings and theories. In: Basic processes in reading: Visual short-term retention. Journal of Verbal Learning and Verbal Behavior
word recognition, ed. D. Besner & G. W. Humphreys. Erlbaum. [SM] 10:34654. [aNC]
Neumann, O. (1987) Beyond capacity: A functional view of attention. In: Peterson, L. R. & Peterson, M. J. (1959) Short-term retention of individual verbal
Perspectives on perception and action, ed. H. Heuer & A. Sanders. Erlbaum. items. Journal of Experimental Psychology 58:19398. [aNC, PM]
[TS] Peterson, S. A. & Simon, T. J. (2000) Computational evidence for the subitizing
Newman, J. B., Baars, B. J. & Cho, S.-B. (1997) A neural Global Workspace model phenomenon as an emergent property of the human cognitive architecture.
for conscious attention. Neural Networks 10:11951206. [BJB] Cognitive Science 24:93122. [NAT, rNC]
Newport, E. L. (1988) Constraints on learning and their role in language Petrides, M. (1996) Lateral frontal cortical contribution to memory. Seminars in
acquisition: Studies of the acquisition of the American Sign Language. the Neurosciences 8:5763. [BR]
Language Sciences 10:14772. [EMP] Pinker, S. (1979) Formal models of language learning. Cognition 7:21783. [EMP]
(1990) Maturational constraints on language learning. Cognitive Science 14:11 (1994) The language instinct. Penguin Books. [EMP]
28. [aNC, EMP] Platt, J. R. (1964) Strong inference. Science 146:34753. [rNC]
Nicolson, R. & Fawcett, A. (1991) Decomposing working memory: New evidence Pollack, I. (1952) The information of elementary auditory displays. Journal of the
from memory span. Paper presented at the International Conference on Acoustical Society of America 24:74549. [BLB]
Memory, Lancaster, U. K., July 1519. [SM] Pollack, I. & Ficks, L. (1954) Information of elementary multi-dimensional
Nissen, M. J. & Bullemer, P. (1987) Attentional requirements of learning: Evidence auditory displays. Journal of the Acoustical Society of America 26:155 58.
from performance measures. Cognitive Psychology 19:132. [aNC] [BLB]
Norman, D. A. & Bobrow, D. G. (1975) On data-limited and resource-limited Pollack, I., Johnson, I. B. & Knaff, P. R. (1959) Running memory span. Journal of
processes. Cognitive Psychology 7:4464. [rNC] Experimental Psychology 57:13746. [aNC, rNC]
Norman, D. A. & Shallice, T. (1986) Attention to action: Willed and automatic Posner, M. I. (1969) Abstraction and the process of recognition. In: Psychology of
control of behavior. In: Consciousness and self-regulation, vol. 4, ed. R. J. learning and motivation, vol. 3, ed. G. H. Bower & J. T. Spence. Academic
Davidson, G. E. Schwartz & D. Shapiro. Plenum. [TS] Press. [aNC]
Nottebohm, F. (1993) The search for neural mechanisms that define the sensitive Posner, M. L. & Konick, A. W. (1966) On the role of interference in short-term
period for song learning in birds. Netherlands Journal of Zoology 43:193234. retention. Journal of Experimental Psychology 72:22131. [PM]
[DT] Posner, M., Snyder, C. & Davidson, B. (1980) Attention and detection of signals.
Oades, R. D., Dittmann-Balcar, A., Zerbin, D. & Grzella, I. (1997) Impaired Journal of Experimental Psychology: General 109:16074. [aNC]
attention-dependent augmentation of MMN in nonparanoid versus paranoid Postman, L. & Phillips, L. W. (1965) Short-term temporal changes in free recall.
schizophrenic patients: A comparison with obsessive-compulsive disorder and Quarterly Journal of Experimental Psychology 17:13238. [aNC]
healthy subjects. Biological Psychiatry 41:11961210. [RDO] Pothos, E. M. (1998) Aspects of generalisation. Unpublished PhD thesis,
Oades, R. D., Zerbin, D., Dittmann-Balcar, A. & Eggers, C. (1996) Auditory event- University of Oxford. [EMP]
related potential (ERP) and difference-wave topography in schizophrenic Potter, M. C. (1976) Short-term conceptual memory for pictures. Journal of
patients with/without active hallucinations and delusions: A comparison with Experimental Psychology: Human Learning and Memory 2:509 22. [GFW]
young obsessive-compulsive disorder (OCD) and healthy subjects. Poulton, E. C. (1954) Eye-hand span in simple serial tasks. Journal of Experimental
International Journal of Psychophysiology 22:185214. [RDO] Psychology 47:40310. [aNC]
Oberauer, K. (in press) Removing irrelevant information from working memory: A Prabhakaran, V., Narayanan, K., Zhao, Z. & Gabrieli, J. D. E. (2000) Integration of
cognitive aging study with the modified Sternberg task. Experimental diverse information in working memory within the frontal lobe. Nature-
Psychology: Learning, Memory and Cognition. [rNC] Neuroscience 3:8590. [BR]
Okada, R. & Burrows, D. (1978) The effects of subsidiary tasks on memory Pylyshyn, Z. (1989) The role of location indices in spatial perception: A sketch of
retrieval from long and short lists. The Quarterly Journal of Experimental the FINST spatial-index model. Cognition 32:6597. [GD]
Psychology 30:22133. [JJ] Pylyshyn, Z., Burkell, J., Fisher, B., Sears, C., Schmidt, W. & Trick, L. (1994)
Oram, M. W. & Perrett, D. I. (1992) Time course of neural responses Multiple parallel access in visual attention. Canadian Journal of Experimental
discriminating different views of the face and head. Journal of Psychology 48:26083. [aNC]
Neurophysiology 68:7084. [PRR] Pylyshyn, Z. W. & Storm, R. W. (1988) Tracking multiple independent targets:
Paivio, A., Yuille, J. C. & Madigan, S. A. (1969) Concreteness, imagery, and Evidence for a parallel tracking mechanism. Spatial Vision 3:179 97. [aNC,
meaningfulness values for 925 nouns. Journal of Experimental Psychology BJS]
Monograph Supplement 76:125. [JSN] Raaijmakers, J. G. W. & Shiffrin, R. M. (1981) Search of associative memory.
Pascual-Leone, J. (1970) A mathematical model for the transition rule in Piagets Psychological Review 88:93134. [aNC]
developmental stages. Acta Psychologica 63:30145. [SM, JPL] Rafal, R. D. (1997) Balint syndrome. In: Behavioral neurology and
(1987) Organismic processes for neo-Piagetian theories: A dialectical causal neuropsychology, ed. T. Feinberg & M. Farrah. McGraw-Hill. [BJS]
account of cognitive development. International Journal of Psychology Raffone, A. & Wolters, G. (in press) A cortical mechanism for binding in visual
22:53170. [SM, JPL] working memory. Journal of Cognitive Neuroscience. [AR]
(1995) Learning and development as dialectical factors in cognitive growth. Raghavachari, S., Caplan, J. B., Kirschen, M., Kahana, M., Madsen, J. R. &
Human Development 38:33848. [JPL] Lisman, J. E. (1999) The Sternberg task evokes theta oscillations in human
(1998) To appraise developmental difficulty or mental demand, relational intracranial recordings. Society for Neuroscience Abstracts 25:1143. [OJ]
complexity is not enough. Behavioral and Brain Sciences 12:84344. [JPL] Rainer, G., Asaad, W. F. & Miller, E. K. (1998) Selective representation of relevant

information by neurons in the primate prefrontal cortex. Nature 393:57779. Scholl, B. J., Pylyshyn, Z. W. & Feldman, J. (in press) What is a visual object?
[PRR] Evidence from multiple object tracking. Cognition. [BJS]
Ratcliff, R. (1978) A theory of memory retrieval. Psychological Review 85:59108. Schubert, T. (1999) Processing differences between choice and simple reaction
[rNC] tasks affect outcome of bottleneck localisation experiments. Journal of
Reber, P. J. & Kotovsky, K. (1997) Implicit learning in problem solving: The role of Experimental Psychology: Human Perception and Performance 25:408 25.
working memory capacity. Journal of Experimental Psychology: General [TS]
126:178 203. [aNC] Schumacher, E., Seymour, T. L., Glass, J. M., Lauber, E. J., Kieras, D. E. & Meyer,
Reisberg, D., Rappaport, I. & OShaughnessy, M. (1984) Limits of working D. E. (1998) Virtually perfect timesharing in dual-task performance:
memory: The digit-digit span. Journal of Experimental Psychology: Learning, Behavioural evidence for independent parallel processing in the human brain.
Memory, and Cognition 10:20321. [aNC] Journal of Cognitive Neuroscience supplement:106. [TS]
Rensink, R. A. (1999) The magical number one, plus or minus zero. Investigative Schweickert, R. & Boruff, B. (1986) Short-term memory capacity: Magic number
Ophthalmology and Visual Science 40:52. [RAR] or magic spell? Journal of Experimental Psychology: Learning, Memory, and
(2000a) Visual search for change: A probe into the nature of attentional Cognition 12:41925. [aNC, SM, JSN]
processing. Visual Cognition 7:345 76. [RAR, BJS] Schweickert, R., Hayt, C., Hersberger, L. & Guentert, L. (1996) How many words
(2000b) Seeing, sensing, and scrutinizing. Vision Research 40:1469 87. [RAR] can working memory hold? A model and a method. In: Models of short-term
(2000c) Differential grouping of features. Perception 29 supplement. [RAR] memory, ed. S. E. Gathercole. Psychology Press. [aNC]
Rensink, R. A., Deubel, H. & Scheneider, W. X. (2000) The incredible shrinking Schyns, P., Goldstone, R. L. & Thilbaut, J.-P. (1998) The development of features
span: Estimates of memory capacity depend on interstimulus interval. in object concepts. Behavioral and Brain Sciences 21:154. [RDO]
Investigative Ophthalmology and Visual Science 41:425. [RAR] Sebrechts, M. M., Marsh, R. L. & Seamon, J. C. (1989) Secondary memory and
Rensink, R. A., ORegan, J. K. & Clark, J. J. (1997) To see or not to see: The need very rapid forgetting. Memory and Cognition 17:693700. [PM]
for attention to perceive changes in scenes. Psychological Science 8:36873. Service, E. (1998) The effect of word length on immediate serial recall depends on
[aNC, RAR] phonological complexity, not articulatory duration. Quarterly Journal of
Richman, H. B., Gobet, F., Staszewski, J. J. & Simon, H. A. (1996) Perceptual and Experimental Psychology 51A:283304. [aNC, SM]
memory processes in the acquisition of expert performance: The EPAM Shah, P. & Miyake, A. (1996) The separability of working memory resources for
model. In: The road to excellence, ed. K. A. Ericsson. Erlbaum. [PCL] spatial thinking and language processing: An individual differences approach.
Richman, H. B., Staszewski, J. J. & Simon, H. A. (1995) Simulation of expert Journal of Experimental Psychology: General 125:427. [aNC]
memory using EPAM IV. Psychological Review 102:30530. [aNC, KAE] Shallice, T. & Warrington, E. K. (1970) Independent functioning of verbal memory
Robertson, L., Treisman, A., Friedman-Hill, S. & Grabowecky, M. (1997) The stores: A neuropsychological study. Quarterly Journal of Experimental
interaction of spatial and object pathways: Evidence from Balints syndrome. Psychology 22:26173. [BJS]
Journal of Cognitive Neuroscience 9:295317. [BJS] Shastri, L. & Ajjanagadde, V. (1993) From simple associations to systematic
Robin, N. & Holyoak, K. J. (1995) Relational complexity and the functions of reasoning: A connectionist representation of rules, variables, and dynamic
prefrontal cortex. In: The cognitive neurosciences, ed. M. S. Gazzaniga. MIT bindings using temporal synchrony. Behavioral and Brain Sciences 16:417 94.
Press. [ JPL] [aNC, HT]
Rodriquez, E., George, N., Lachaux, J.-P., Martinerie, J., Renault, B. & Varela, F. J. Shibuya, H. & Bundesen, C. (1988) Visual selection from multielement displays:
(1999) Perceptions shadow: long-distance synchronization of human brain Measuring and modeling effects of exposure duration. Journal of Experimental
activity. Nature 397:43033. [aNC] Psychology: Human Perception and Performance 14:16068. [WS]
Roelfsema, P. R., Lamme, V. A. F. & Spekreijse, H. (2000) The implementation of Shiffrin, R. M. (1993) Short-term memory: A brief commentary. Memory and
visual routines. Vision Research 40:13851411. [PRR] Cognition 21:19397. [aNC, rNC]
Roelfsema, P. R. & Singer, W. (1998) Detecting connectedness. Cerebral Cortex Shiffrin, R. M. & Schneider, W. (1977) Controlled and automatic human
8:385 96. [PRR] information processing: II. Perceptual learning, automatic attending, and a
Roitblat, H. L., Bever, T. G., Helweg, D. A. & Harley, H. E. (1991) On-line choice general theory. Psychological Review 84:12790. [aNC, rNC]
and the representation of serially structured stimuli. Journal of Experimental Simon, H. A. (1974) How big is a chunk? Science 183:48288. [aNC]
Psychology: Animal Behavior Processes 17:5567. [DT] Simon, T. J. & Vaishnavi, S. (1996) Subitizing and counting depend on different
Rubin, D. C. & Wenzel, A. E. (1996) One hundred years of forgetting: A quantitative attentional mechanisms: Evidence from visual enumeration in afterimages.
description of retention. Psychological Review 103:73460. [PM] Perception and Psychophysics 58:91526. [aNC, BJS]
Ryan, J. (1969) Grouping and short-term memory: Different means and patterns of Simons, D. J. (1996) In sight, out of mind: When object representations fail.
groups. Quarterly Journal of Experimental Psychology 21:13747. [aNC] Psychological Science 7:301 305. [RAR]
Rypma, B., Prabhakaran, V., Desmond, J. E., Glover, G. H. & Gabrieli, J. D. E. Simons, D. J. & Levin, D. T. (1998) Failure to detect changes to people during a
(1999) Load-dependent roles of prefrontal cortical regions in the maintenance real-world interaction. Psychonomic Bulletin and Review 5:64449. [aNC]
of working memory. NeuroImage 9:21626. [BR] Sirevaag, E. J., Kramer, A. F., Coles, M. G. H. & Donchin, E. (1989) Resource
Sanders, A. F. (1968) Short term memory for spatial positions. Psychologie 23:1 reciprocity: An event-related brain potentials analysis. Acta Psychologica
15. [aNC] 70:7797. [aNC]
Sanders, A. F. & Schroots, J. J. F. (1969) Cognitive categories and memory span: Slak, S. (1970) Phonemic recoding of digital information. Journal of Experimental
III. Effects of similarity on recall. Quarterly Journal of Experimental Psychology 86:398406. [DJM]
Psychology 21:2128. [aNC] Slemecka, N. J. (1985) Ebbinghaus: Some associations. Journal of Experimental
Saults, J. S. & Cowan, N. (1996) The development of memory for ignored speech. Psychology: Learning, Memory, and Cognition 11:41435. [JJ, rNC]
Journal of Experimental Child Psychology 63:23961. [rNC] Smith, E. E. & Jonides, J. (1995) Working memory in humans: Neuropsychological
Scarborough, D. L. (1971) Memory for brief visual displays: The role of implicit Evidence. In: The cognitive neurosciences, ed. M. Gazzaniga. MIT Press.
speech. Paper presented to the Eastern Psychological Association, New York, [ADB]
NY. [aNC] (1999) Storage and executive processes in the frontal lobes. Science 283:1657
Schneider, W. X. (1999) Visual-spatial working memory, attention, and scene 61. [BR]
representation: A neuro-cognitive theory. Psychological Research 62:22036. Solso, R. L. (1995) Cognitive Psychology, 4th ed. Allyn and Bacon. [PM]
[WS] Sperling, G. (1960) The information available in brief visual presentations.
Schneider, W. & Detweiler, M. (1987) A connectionist/control architecture for Psychological Monographs 74:Whole No. 498. [aNC, SAH, RDO, TS, rNC]
working memory. In: The psychology of learning motivation, vol. 21, ed. G. H. (1967) Successive approximations to a model for short-term memory. Acta
Bower. Academic Press. [aNC] Psychologica 27:28592. [aNC]
Schneider, W. & Shiffrin, R. M. (1977) Controlled and automatic human Spitz, H. H. (1973) The channel capacity of educable mental retardates. In: The
information processing: I. Detection, search, and attention. Psychological experimental psychology of mental retardation, ed. D. K. Routh. Aldine.
Review 84:1 66. [aNC, rNC] [BLB]
Schneider, W. X., Wesenick, M. B., Deubal, H. & Bundesen, C. (1999) A study of Spitzer, M. (1997) A cognitive neuroscience view of schizophrenic thought
visuospatial working memory. Perception 28:5. [RAR] disorder. Schizophrenia Bulletin 23:2950. [RDO]
Scholl, B. J. (in press) Objects and attention: The state of the art. Cognition: Stadler, M. (1989) On learning complex procedural knowledge. Journal of
Special Issue on Objects and Attention. [BJS] Experimental Psychology: Learning, Memory, and Cognition 15:1061 69.
Scholl, B. J. & Leslie, A. M. (1999) Explaining the infants object concept: Beyond [aNC]
the perception/cognition dichotomy. In: What is cognitive science?, ed. E. Steele, C. M. & Josephs, R. A. (1990) Alcohol myopia: Its prized and dangerous
Lepore & Z. Pylyshyn. Blackwell. [BJS] effects. American Psychologist 45:92133. [rNC]
Scholl, B. J. & Pylyshyn, Z. W. (1999) Tracking multiple items through occlusion: Steriade, M., McCormick, D. A. & Sejnowski, R. J. (1993) Thalamocortical
Clues to visual objecthood. Cognitive Psychology 38:25990. [BJS] oscillations in the sleeping and aroused brain. Science 262:679685. [BJB]

Sternberg, S. (1966) High-speed scanning in human memory. Science 153:65254. word frequency on immediate recall and learning of verbal material.
[aNC, BJB, GG, JJ, DJM, rNC] Canadian Journal of Psychology 16:8395. [aNC]
(1969a) The discovery of processing stages: Extensions of Donders method. Tulving, E. & Patterson, R. D. (1968) Functional units and retrieval processes in
Acta Psychologica 30:276315. [BR] free recall. Journal of Experimental Psychology 77:23948. [aNC]
(1969b) Memory scanning: Mental processes revealed reaction-time Tulving, E. & Pearlstone, Z. (1966) Availability versus accessibility of information
experiments. American Scientist 57:42157. [JJ] in memory for words. Journal of Verbal Learning and Verbal Behavior 5:381
(1975) Memory scanning: New findings and current controversies. Quarterly 91. [aNC]
Journal of Experimental Psychology 27:132. [JJ] Turvey, M. T. (1973) On peripheral and central processes in vision: Inferences
Stewart, L. & Pascual-Leone, J. (1992) Mental capacity constraints and the from an information processing analysis of masking with patterned stimuli.
development of moral reasoning. Journal of Experimental Child Psychology Psychological Review 80:152. [rNC]
54:251 87. [ JPL] Uller, C., Carey, S., Huntley-Fenner, G. & Klatt, L. (1999) What representations
Stigler, J. W. (1984) Mental abacus: The effect of abacus training on Chinese might underlie infant numerical knowledge? Cognitive Development 14:1 43.
childrens mental calculation. Cognitive Psychology 16:14576. [KAE] [BJS]
Straube, E. & Oades, R. D. (1992) Schizophrenia: Empirical research and finding. Usher, M. & Cohen, J. (1999) Short term memory and selection processes in a
Academic Press. [RDO] frontal-lobe model. In: Connectionist models in cognitive neuroscience, ed. D.
Sugase, Y., Yamane, S., Ueno, S. & Kawano, K. (1999) Global and fine information Heinke, G. W. Humphries & A. Olsen. Springer-Verlag. [MU]
encoded by single neurons in the temporal visual cortex. Nature 400:86972. Usher, M. & McClelland, J. L. (1995) Time course of perpetual choice. (Tech. Rep.
[RDO] No. PDP. CNS.95.5). Pittsburgh, PA: Carnegie Mellon University. [MU]
Taatgen, N. A. (1999a) Cognitief modelleren, een nieuwe kijk op individuele Vallar, G. & Baddeley, A. D. (1982) Short-term forgetting and the articulatory loop.
verschillen. Nederlands tijdschrift voor de psychologie 54:16777. [NAT] Quarterly Journal of Experimental Psychology 34A:5360. [aNC]
(1999b) A model of learning task-specific knowledge for a new task. In: (1984) Fractionation of working memory: Neuropsychological evidence for a
Proceedings of the Twenty-first Annual Meeting of the Cognitive Science phonological short-term store. Journal of Verbal Learning and Verbal
Society. Erlbaum. [NAT] Behavior 23:15161. [BR]
Tallon-Baudry, C., Kreiter, A. & Bertrand, O. (1999) Sustained and transient Van Oeffelen, M. & Vos, P. (1982) A probabilistic model for the discrimination of
oscillatory responses in the gamma and beta bands in a visual short-term visual number. Perception and Psychophysics 32:16370. [SEA]
memory task in humans. Visual Neuroscience 16:44959. [OJ] Van Strien, J. W. & Bouma, A. (1990) Selective activation effects on concurrent
Tan, L. & Ward, G. (in press) A recency-based account of the primacy effect in verbal and spatial memory loads in left-handed and right-handed adults. Brain
free recall. Journal of Experimental Psychology: Learning, Memory and and Cognition 14:8191. [SM]
Cognition. [SEA] Varela, F. J. (1995) Resonant cell assemblies: A new approach to cognitive
Tehan, G. & Humphreys, M. S. (1996) Cuing effects in short-term recall. Memory functions and neuronal synchrony. Biological Research 28:81 95. [AR]
and Cognition 24:71932. [aNC, JSN, rNC] Vogel, E. K. (2000) Selective storage in visual working memory: Distinguishing
Terrace, H. (1987) Chunking by a pigeon in a serial learning task. Nature 325:149 between perceptual-level and working memory-level mechanisms of attention.
51. [DT] University of Iowa Press. [GFW]
Thompson, C. P., Cowan, T. M. & Frieman, J. (1993) Memory search by a Vogel, E. K., Luck, S. J. & Shapiro, K. L. (1998) Electrophysiological evidence for
memorist. Erlbaum. [DJM, JW] a postperceptual locus of suppression during the attentional blink. Journal of
Tiitinen, H. & May, P. (preprint) Auditory transient and sustained responses as a Experimental Psychology: Human Perception and Performance 24:1656 74.
function of interstimulus interval. [HT] [aNC, GFW]
Tiitinen, H., May, P., Reinikainen, K. & Ntnen, R. (1994) Attentive novelty Von der Malsburg, C. (1981) The correlation theory of brain function. (Internal
detection in humans is governed by pre-attentive sensory memory. Nature Report 812). Gttingen, NL: Max Planck Institute for Biophysical
372:90 92. [HT] Chemistry. [AR]
Tiitinen, H., Sinkkonen, J., Reinikainen, K., Alho, K., Lavikainen, J. & Ntnen, (1995) Binding in models of perception and brain function. Current Opinion in
R. (1993) Selective attention enhances the auditory 40-Hz transient response Neurobiology 5:52026. [aNC, PRR]
in humans. Nature 364:5960. [aNC, HT] Warren, R. M. & Obusek, C. J. (1972) Identification of temporal order within
Todt, D., Cirillo, J., Geberzahn, N. & Schleuss, F. (2000) The role of hierarchy auditory sequences. Perception and Psychophysics 12:8690. [CPB, rNC]
levels in vocal imitations of songbirds. Cybernetics and Systems: An Warren, R. M., Obusek, C. J., Farmer, R. M. & Warren, R. P. (1969) Auditory
International Journal 32:127. [DT] sequence: Confusion of patterns other than speech or music. Science
Tomonaga, M. & Matsuzawa, T. (2000) Sequential responding to arabic numbers 164:58687. [CPB]
with wild cards by the chimpanzee (Pan troglodytes). Animal Cognition 3:1 Watkins, M. J. (1974) Concept and measurement of primary memory.
11. [NK] Psychological Bulletin 81:695711. [aNC]
Toms, M., Morris, N. & Ward, D. (1993) Working memory and conditional Watkins, M. J. & Watkins, O. C. (1974) Processing of recency items for free recall.
reasoning. Quarterly Journal of Experimental Psychology 46A:67999. Journal of Experimental Psychology 102:48893. [PM]
[aNC] Watkins, O. C. & Watkins, M. J. (1975) Build-up of proactive inhibition as a cue-
Tononi, G. & Edelman, G. M. (1998) Consciousness and complexity. Science overload effect. Journal of Experimental Psychology: Human Learning and
282:1846 51. [BJB] Memory 1:44252. [aNC]
Townsend, J. T. (1976) Serial and within-stage independent parallel model Watson, S. E. & Kramer, A. F. (1999) Object-based visual selective attention and
equivalence on the minimum completion time. Journal of Mathematical perceptual organization. Perception and Psychophysics 61:31 49. [GD]
Psychology 14:21938. [rNC] Waugh, N. C. & Norman, D. A. (1965) Primary memory. Psychological Review
Towse, J. N. & Houston-Price, C. M. T. (in press) Reflections on the concept of the 72:89104. [aNC, JJ, BR]
central executive. In: Working memory in perspective, ed. J. Andrade. Wegner, D. M., Quillian, F. & Houston, C. E. (1996) Memories out of order:
Psychology Press. [JNT] Thought suppression and the disturbance of sequence memory. Journal of
Treisman, A. M. & Gelade, G. (1980) A feature integration theory of attention. Personality and Social Psychology 71:68091. [CPB]
Cognitive Psychology 12:97136. [PRR, rNC] Welford, A. T. (1980) The single-channel hypothesis. In: Reaction times, ed. A. T.
Treisman, A. & Gormican, S. (1988) Feature analysis in early vision: Evidence from Welford. Academic Press. [TS]
search asymmetries. Psychological Review 95:1548. [WS] Wharton, R. M. (1974) Approximate language identification. Information and
Trick, L. M. & Pylyshyn, Z. W. (1993) What enumeration studies can show us Control 26:23655. [EMP]
about spatial attention: Evidence for limited capacity preattentive processing. Wickelgren, W. A. (1964) Size of rehearsal group and short-term memory. Journal
Journal of Experimental Psychology: Human Perception and Performance of Experimental Psychology 68:41319. [aNC]
19:331 51. [aNC, GD, TS] (1966) Phonemic similarity and interference in short-term memory for single
(1994a) Why are small and large numbers enumerated differently? A limited- letters. Journal of Experimental Psychology 71:396404. [aNC]
capacity preattentive stage in vision. Psychological Review 101:80102. Wickelgren, W. A., Corbett, A. T. & Dosher, B. A. (1980) Priming and retrieval
[aNC, SEA, BJS] from short-term memory: A speed-accuracy tradeoff analysis. Journal of
(1994b) Cueing and counting: Does the position of the attentional focus affect Verbal Learning and Verbal Behavior 19:387404. [BM]
enumeration? Visual Cognition 1:67100. [aNC] Wickelgren, W. A. & Norman, D. A. (1966) Strength models and serial position in
Tsodyks, M., Pawelzik, K. & Markram, H. (1998) Neural networks with dynamic short-term memory. Journal of Mathematical Psychology 3:316 47.
synapses. Neural Computation 10:82135. [MU] [DJM]
Tulving, E. & Colotla, V. (1970) Free recall of trilingual lists. Cognitive Psychology Wickens, C. D. (1984) Processing resources in attention. In: Varieties of attention,
1:86 98. [aNC] ed R. Parasuraman & D. R. Davies. Academic Press. [aNC]
Tulving, E. & Patkau, J. E. (1962) Concurrent effects of contextual constraint and Wickens, D. D., Moody, M. J. & Dow, R. (1981) The nature and timing of the

retrieval process and of interference effects. Journal of Experimental top-down allocation of attention to global and local features. Journal of
Psychology: General 110:120. [aNC] Neuroscience 20:15. [RDO]
Wilkes, A. L. (1975) Encoding processes and pausing behaviour. In: Studies in Yantis, S. (1992) Multi-element visual tracking: Attention and perceptual
long-term memory, ed. A. Kennedy & A. Wilkes. Wiley. [aNC] organization. Cognitive Psychology 24:295340. [aNC, SEA]
Wilson, B. A. & Baddeley, A. D. (1988) Semantic episodic and autobiographical Yantis, S. & Johnson, D. (1990) Mechanisms of attentional priority. Journal of
memory in a postmeningitic amnesic patient. Brain and Cognition 8:3146. Experimental Psychology: Human Perception and Performance 16:81225.
[ADB] [BJS]
Woodman, G. F. & Luck, S. J. (1999) Electrophysiological measurement of rapid Young, R. M. (1996) Functionality matters: Capacity constraints and Soar. In: Mind
shifts of attention during visual search. Nature 400:86769. [GFW] matters: A tribute to Allen Newell, ed. D. M. Steier & T. M. Mitchell.
Woodman, G. F., Vogel, E. K. & Luck, S. J. (in press) Visual search remains Erlbaum. [CPB]
efficient when visual working memory is full. Psychological Science. [GFW, Young, R. M. & Lewis, R. L. (1999) The Soar cognitive architecture and human
rNC] working memory. In: Models of working memory: Mechanisms of active
Wynn, K. (1992) Addition and subtraction by human infants. Nature 358:74950. maintenance and executive control, ed. A. Miyake & P. Shah. Cambridge
[BJS] University Press. [CPB]
Xu, Y. (submitted) Integrating color and shape in visual short-term memory for Zhang, G. & Simon, H. A. (1985) STM capacity for Chinese words and idioms:
objects with parts. [BJS] Chunking and acoustical loop hypotheses. Memory and Cognition 13:193
Yamaguchi, S., Yamagata, S. & Kobayashi, S. (2000) Cerebral asymmetry of the 201. [aNC, SEA]

Cowan BBS 2001

Uploaded by

Copyright:

Available Formats

Cowan BBS 2001

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Cowan BBS 2001

Uploaded by

Copyright:

Available Formats

BEHAVIORAL AND BRAIN SCIENCES (2000) 24, 87185

Printed in the United States of America

The magical number 4 in short-term

2000 Cambridge University Press 0140-525X/00 $12.50 87

88 BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1 89

3.1. Imposing an information overload

90 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1 91

92 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1 93

94 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1 95

96 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1 97

98 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1 99

Reference Data Source Method of Calculating Items in Storage Est.

100 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1 101

102 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1 103

104 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1 105

106 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1 107

108 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1 109

110 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1 111

112 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1 113

114 BEHAVIORAL AND BRAIN SCIENCES (2001) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1 115

116 BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1 117

118 BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1 119

120 BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1 121

122 BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1 123

124 BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1 125

126 BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1 127

128 BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1 129

130 BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1 131

132 BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1

Equation 1 can be written alliteratively as

As L tends to infinity, it can be seen from Equation 2 that x tends

BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1 133

134 BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1

Third, and most relevant to the remainder of our discussion,

BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1 135

136 BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1

BEHAVIORAL AND BRAIN SCIENCES (2000) 24:1 137