Consider the earlier example of short-term memory loss: of entering a room and forgetting your reason for doing so. Also consider distortions of perception

R E V I E W S

NAT U R E R EV I EWS | N E U RO S C I E N C E VO LU M E 4 | O C TO B E R 2 0 0 3 | 8 2 9

The theoretical concept of working memory assumes that a limited capacity system, which temporarily maintains and stores information, supports human thought processes by providing an interface between perception, long-term memory and action1–3. There are many approaches to the study of working mem- ory, using a range of empirical and theoretical tech- niques. However, most theories agree on the need for a system of limited attentional capacity, supple- mented by more peripherally-based storage systems4. The account that follows emphasizes this fractiona- tion, stressing the importance of executive control, but concentrating initially on the more tractable peripheral storage systems. Within experimental cog- nitive psychology there are several different but com- plementary approaches to working memory: some emphasize the role of attentional control in memory5, whereas others attempt to explain working memory data in terms of models that were originally devel- oped for long-term memory (LTM)6,7. An example of this approach is provided by Nairne8, whose account seems to criticize the multi-component model of working memory. However, most of Nairne’s criticisms apply only if one attempts to explain all the phenomena of working memory in terms of the phonological loop — one component of a complex system. As this review shows, the phonological loop has received more attention than other components, but this reflects its relative tractability, compared with the visuospatial and executive subsystems. A third,

influential approach relies on correlational techniques that capitalize on individual differences across the various components of working memory (BOX 1). Rather than attempting to give an account of each of these approaches, I describe a single multi-component model of working memory. In general, deviations from the other models represent difference of emphasis and scope, rather than direct conflict. Points of clear disagreement will, however, be identified and discussed.

The term ‘working memory’ seems to have been invented by Miller, Galanter and Pribram9, and was adopted by Baddeley and Hitch10 to emphasize the differences between their three-component model and earlier unitary models of short-term memory (STM). These differences include its multi-component character, its emphasis on combined processing and storage, and the stress on its functional importance as a system that facilitates a range of cognitive activities, such as reasoning, learning and comprehension. This approach resulted in the development of a set of experimental tasks that could be used to analyse different activities and subject populations1. Because the available empirical evidence provided few con- straints, the model was initially loosely specified: for example, it had no mechanism for storing serial order. However, it was simple and robust, and had the potential to develop and become much richer and more clearly specified. The Baddeley and Hitch model continues to flourish, and will be used as a basis for this review. The article will, however, extend beyond

WORKING MEMORY: LOOKING BACK AND LOOKING FORWARD Alan Baddeley

The concept of working memory proposes that a dedicated system maintains and stores information in the short term, and that this system underlies human thought processes. Current views of working memory involve a central executive and two storage systems: the phonological loop and the visuospatial sketchpad. Although this basic model was first proposed 30 years ago, it has continued to develop and to stimulate research and debate. The model and the most recent results are reviewed in this article.

Department of Psychology, University of York, York YO10 5DD, UK. e-mail: [email protected] doi:10.1038/nrn1201

830 | O C TO B E R 2 0 0 3 | VO LU M E 4 www.nature.com/reviews/neuro

R E V I E W S

memory, then patients with STS impairment should show little capacity for long-term learning or for every- day cognitive activities. Such patients were identified18, but had few cognitive problems beyond grossly impaired STM19.

Baddeley and Hitch10 used secondary tasks to deplete the availability of STM in subjects performing tasks, such as reasoning or learning, that were assumed to rely on working memory. They found clear but far from catastrophic impairment, and proposed a three- component model of working memory in place of the unitary system. The three components (FIG. 1) comprised a control system of limited attentional capacity, termed the central executive, which is assisted by two subsidiary storage systems: the phonological loop, which is based on sound and language, and the visuospatial sketchpad.

In this review, I focus on each of these components in turn, emphasizing in particular those aspects that have developed or changed in recent years.

The phonological loop Structure of the loop. The model of the phonological loop comprises a phonological store, which can hold memory traces for a few seconds before they fade, and an articulatory rehearsal process that is analogous to subvocal speech (FIG. 2). Memory traces can be refreshed by being retrieved and re-articulated. Immediate mem- ory has a limited span because articulation takes place in real time — as the number of items rehearsed increases, it reaches a point at which the first item will have faded before it can be rehearsed. Much of the work in this area has used immediate serial recall, typically using a small set of digits, letters or unrelated words, with the characteristics of the material remembered being used to give an indication of the nature of the code on which the recall is based. For unrelated letters, the code is acoustic or phonological; sequences of simi- lar sounding letters such as V, B, G, T, P, C are recalled less well than a dissimilar set, such as W, X, K, R, Y, Q20,21. Similarity of sound is also crucial for unrelated words, whereas meaning is relatively unimportant22. When the model is switched from STM to LTM using several presentations of longer lists, sound becomes irrelevant and meaning crucial23.

Evidence for the role of articulation comes from the word-length effect: immediate memory span declines as word length increases from one to five syllables24. The suggestion that this reflects the slower rehearsal of

the range of the model to speculations as to the proba- ble future development of a more general theory of working memory.

The multi-component model of working memory In 1949, Hebb11 proposed a distinction between STM, which is based on temporary electrical activation, and LTM, which is based on neuronal growth. A decade later, support for this distinction came from studies showing that small amounts of information were rapidly forgotten unless actively rehearsed12,13. A counter attack14 argued that results such as these could be explained in terms of a unitary LTM system, but in the mid-1960s many studies argued for a separation between LTM and STM15. By the late 1960s, many new models had emerged around the concept of STM. The most influential model16 proposed that information from the environment flows through a series of tempo- rary sensory registers into a limited capacity short-term store (STS), which feeds information into and out of LTM. This system was also assumed to act as a working memory, supporting complex cognitive activities, although this aspect was not further explored.

This model encountered problems both in terms of its assumptions regarding learning17, and from data concerned with the impact of neuropsychological dam- age to the STS. If the STS served as a unitary working

Box 1 | Individual differences in working memory

In 1980, Daneman and Carpenter164 developed a task in which subjects were required to combine storage and processing, first reading a series of unrelated sentences, and then recalling the final word of each. Working memory span was defined as the maximum number of sentences for which this task could be performed perfectly. They found a high correlation between working memory span and reading comprehension, a result that has been replicated many times165. Similar results occur when sentence processing is replaced by other tasks, such as arithmetic calculation166 or colour–word association167. Prediction is not limited to comprehension but extends from spelling168 to acquisition of logic169, and from note taking170 to following directions171, with correlations typically being in the region of 0.5. Kyllonen and Christal172 compared a cluster of working memory span measures with more conventional measures of intelligence based on reasoning. They found a high correlation, with the main difference being a slightly higher reliance of working memory span on processing speed, and of reasoning on previous knowledge.

Although much of the work in this area has simply assessed the correlation between working memory span and performance on cognitive tasks, there is a growing interest in understanding the underlying processes that contribute to the relatively complex working memory span task. Studies using factor analysis and latent variable analysis have supported the broad concept of separate phonological173 and visuospatial storage systems174, together with a main executive system. Attention is now focusing on the processes that underlie executive control173,175,176.

Theories of intelligence have typically been developed as statistical models of data from large batteries of tests. These methods have been used to address controversies such as whether intelligence is best regarded as a single general capacity (G), as proposed by Spearman177, or as a range of separate capacities, as proposed by Thurstone178. A meta- analysis by Carroll179 of more than 400 studies has identified three components, one resembling Spearman’s G, one visuospatial and one verbal-arithmetic. The analogy with the three-component working memory model is clear. Although extremely successful practically, the psychometric study of intelligence seems, in recent years, to have generated more political controversy than theoretical progress180,181. Applying the sophisticated statistical methods that have been developed within psychometrics to the experimentally and neuropsychologically grounded multi-component working memory model might be a fruitful way forward.

Visuospatial sketchpad

Central executive

Phonological loop

Figure 1 | The three-component model of working memory. This model, proposed by Baddeley and Hitch10, comprises a control system, the central executive, and two storage systems, the visuospatial sketchpad and the phonological loop.

NAT U R E R EV I EWS | N E U RO S C I E N C E VO LU M E 4 | O C TO B E R 2 0 0 3 | 8 3 1

R E V I E W S

interpretations assume that each item forms a cue or stimulus for the following item, with the result that once the initial item is activated, the sequence runs off rela- tively automatically31,32. Such models have difficulty in dealing with sequences in which the same item recurs (7, 1, 9, 3, 1, 5, 8), which leads to a slight impairment in performance33. More problematic for chaining models is the pattern of data that is observed when phonologi- cally similar and dissimilar items are interleaved (C, X, P, W, D, Y). According to chaining models, similarity should cause confusion among similar stimuli, leading to the cueing of the wrong responses. Errors should therefore follow similar letters and involve the dissimilar items. In fact, performance on dissimilar items is unaffected by the interleaving of similar items, which continue to show a higher error rate30,34.

A range of computationally explicit models of ver- bal STM have been proposed6,7,31, with several based on the phonological loop concept35–38. Burgess and Hitch36 assume that order is carried by associating successive items with an ongoing contextual cue, the exact nature of which is unspecified. The phonologi- cal similarity effect occurs because the items that are specified by each cue are encoded phonologically, with similar items having fewer distinguishing cues. The similarity effect is assumed to occur at retrieval rather than at the cueing stage.

The Page and Norris model38 assumes that recall of order is based on positional associations between the first and subsequent items. The associations become progressively weaker as more items are added, providing a simple explanation for limited memory span. Retrieval involves competitive queueing: the strongest association is retrieved first and the associated item is emitted and then inhibited, allowing the next strongest to be retrieved and so on to the end of the list, or to the point at which the associations become too weak and the process breaks down. A related model that also uses serial position as its context is Henson’s start–end model37, in which both the first and last items act as cues.

All of the contextual models give good accounts of the distribution of intrusion errors, whereby the most common error is for two items to transpose (present 1, 2, 3, 4, 5, 6; recall 1, 2, 3, 5, 4, 6). Henson’s model, however, gives a better account of what he calls protrusions, a tendency for an omitted item to be replaced by an item in the same serial position within the previous list (present 1, 2, 3, 4, 5, recall 1, 2, 3, 4, 5; then present 6, 7, 8, 9, 0, recall 6, 7, 3, 9, 0). It seems possible, however, that this reflects a separate, more long-term component, as discussed later.

Despite differences in the nature of the contextual cues, computationally explicit models are possible, and they typically separate the mechanism for storing order from the mechanism by which the items are registered. Bearing this in mind, we should return to the basic phenomena that prompted the model.

The phonological similarity effect. The phonological similarity effect is highly robust39, and is often used as a marker of the phonological loop. It tends to disappear

longer words is supported by the abolition of the word- length effect when subvocal rehearsal is prevented by requiring subjects to repeat an irrelevant sound24,25.

The two-component model of the loop is also supported by neuropsychological evidence (see REF. 19 for a recent review). Patients with phonological STM deficits in the absence of more general language impairment typically have lesions in the left temporoparietal area (BRODMANN AREA (BA) 40). When words are presented visually, these patients show neither a phonological similarity effect nor a word-length effect, a pattern of behaviour that is consistent with their avoiding the use of the defective phonological store, despite having a nor- mal capacity to articulate26. Subvocal rehearsal does not seem to depend on the capacity for peripheral control of articulation, as it can be preserved in ANARTHRIC patients who have lost the capacity for overt articulation27. On the other hand, DYSPRAXIC patients who have lost the capacity to set up internal speech–motor programmes do show the predicted reduction in memory span28.

Although the two-component model of the loop gave a good account of a range of evidence, it was a lim- ited model that left three crucial questions unanswered. How is serial order maintained? How does the loop interact with LTM? And what is the biological function of the phonological loop?

The problem of serial order Explanations of how serial order is remembered, learned and retrieved fall into two broad categories: chaining models and contextual models29,30. Chaining

BRODMANN AREAS

(BA). Korbinian Brodmann (1868–1918) was an anatomist who divided the cerebral cortex into numbered subdivisions on the basis of cell arrangements, types and staining properties (for example, the dorsolateral prefrontal cortex contains subdivisions, including BA 46, BA 9 and others). Modern derivatives of his maps are commonly used as the reference system for discussion of brain- imaging findings.

ANARTHRIC

Unable to speak because of defective articulation.

DYSPRAXIC

Having an impairment of the ability to perform certain voluntary movements, often including speech.

Auditory input

Phonological analysis

Phonological output buffer

Broca’s area–premotor cortex

Spoken output

Rehearsal process

Phonological

STS

Inferior parietal lobe

Visual input

Visual analysis and STS

Orthographic to phonological

recoding

a

b

c

d

e

Figure 2 | A functional model of the phonological loop. a | Phonological analysis. b | Short-term storage (STS). c | The programming of speech output. d | Visual encoding. e | Grapheme-to-phoneme conversion. Auditory information gains direct access to a phonological store, after which it passes to an output buffer for recall or recycling through rehearsal. Visual material can be recoded verbally and then gain access to the phonological store through rehearsal. Modified, with permission, from REF. 19  (2002) John Wiley & Sons Ltd.

832 | O C TO B E R 2 0 0 3 | VO LU M E 4 www.nature.com/reviews/neuro

R E V I E W S

items to be ignored60–62. This prompted alternative inter- pretations of the effect7,41,55,63. Jones and colleagues55,63

have found that the effect is not limited to speech or music, but can also be produced by variable tones. The crucial requirement seems to be a fluctuation in the state of the irrelevant stimulus stream. They suggest that this sets up a competing representation that disrupts the storage of serial order.

The proposal of separate item and order mechanisms in the phonological loop allows a range of explanations of the irrelevant sound effect40. A detailed account within the primacy model of the phonological loop is given by Page and Norris64, but this remains an area of lively controversy41,65. Indeed, Jones and colleagues deny the need to assume separate visual and verbal storage systems65, although the evidence they cite has not proved readily replicable, and their model fails to give a good account of data from neuropsychological and neuroimaging studies40.

Function of the phonological loop. We have proposed that the phonological loop evolved to facilitate the acquisition of language66. Evidence for this view came initially from the failure of a patient with a pure phono- logical loop deficit to acquire the vocabulary of a new language, despite otherwise normal verbal LTM66. It was supported by the demonstration that factors that disrupt the phonological loop — such as articulatory suppression, phonological similarity and word length — also disrupt the acquisition of foreign vocabulary, but not of learning to associate pairs of unrelated native language words; such learning is typically based on semantic coding67,68. Phonological loop capacity is a good predictor of the ability of children69 and adults70 to learn a second language.

Acquisition of native vocabulary in children is well predicted by non-word repetition, the capacity to hear and repeat back an unfamiliar pseudo-word, which is assumed to depend on the phonological loop71,72. Children with a specific language disability, coupled with normal non-verbal intelligence, also perform badly on non-word repetition, but have no obvious hearing or articulatory problems73. This and further evidence led to our proposal that the phonological loop evolved to facilitate language acquisition66.

Most studies of verbal STM rely on the retention of sequences of items such as digits and letters that are already familiar. If we are correct, however, this system has evolved to master new words, comprising unfamiliar sequences of phonemes. Studies within the field are therefore beginning to focus strongly on the sublexical level of analysis, and on the more detailed structure of the phonological loop system74,75. There are two main questions: how temporary storage can enhance the learning of new words, and conversely, how language influences STM.

Immediate recall of non-words is better when they are similar in phonotactic structure to the native language of the person remembering76. This indicates that long-term implicit knowledge can be used to aid immedi- ate recall. However, performance on phonotactically

when error rates increase beyond around 50%, indicating that subjects are abandoning the loop and opting for alternative strategies such as sematic or visual coding 40,41. Given the importance of strategy, there have been surprisingly few attempts to control it by instruction, with the notable exception of Hanley and Bakopoulou42; we are likely to see more such studies, preferably backed up by neuroimaging measures.

The word-length effect. Although almost as robust as the phonological similarity effect39, the word-length effect is open to a wider range of interpretations43. One approach has been to reject time-based decay, arguing that long words are more difficult to recall because they contain more components and are therefore more frag- ile44,45. When long and short words are mixed, however, long words are no more difficult to remember than their shorter neighbours, with recall depending on overall list duration, as predicted by the loop hypothesis46. Finally, the abolition of the word-length effect for either auditory or visual presentation, when rehearsal is suppressed24,47, fits more naturally into the phonological loop model than into its competitors.

Less radical modifications of the model have also been proposed. It has been suggested that the word- length effect stems largely43,46, or even entirely48, from delay during output, rather than from rehearsal. Recent experiments controlling output time, however, implicate both, as the loop model would predict47.

The assumption that forgetting in the phonological store reflects trace decay has also been challenged. Evidence for time-based decay came initially from a study comparing recall of disyllabic words that were spoken quickly (bishop, tipple) or slowly (harpoon, Friday)24. As predicted, longer duration words were less well recalled. However, there were failures to replicate this using other sets of simuli28,49. It is possible that, in these studies, spoken duration was not adequately measured, and the two sets were not equated for phono- logical similarity50 (but see REF. 51), but other studies using different sets of study words further questioned the importance of duration relative to other factors such as linguistic complexity52. However, the most method- ologically rigorous study so far, which considered the whole range of materials that have been used previously, concluded that when duration and similarity are measured carefully, all sets of material give results that can be accounted for in terms of spoken duration and phonological similarity53. It seems, therefore, that the simple trace decay assumption, though controversial, is still adequate, obviating the need for more complex interference models.

Irrelevant sound effects. Immediate recall is impaired by the concurrent or subsequent presentation of irrelevant spoken material7,54–57. This was initially interpreted in terms of mnemonic masking within STM57,58. However, irrelevant speech had no greater effect on phonologi- cally similar items than on dissimilar items59, and per- formance was unaffected by phonological similarity between the material to be remembered and the spoken

NAT U R E R EV I EWS | N E U RO S C I E N C E VO LU M E 4 | O C TO B E R 2 0 0 3 | 8 3 3

R E V I E W S

which a random 50% are filled (FIG. 3b). The matrix is then removed, and the subject attempts to recall which cells were filled. Testing begins with a 2 × 2 matrix, with the matrix size increasing until performance breaks down. Della Sala et al.85 have shown a double dissocia- tion between visual and spatial span. For normal sub- jects, the Corsi task is disrupted more by spatial than visual interference, whereas the reverse is true for pat- tern span (FIG. 3c). Neuropsychological cases have also

unfamiliar items was much better at predicting subse- quent vocabulary acquisition than performance on phonotactically familiar items. It is possible that both types of material depend on the phonological store, which is responsible for long-term learning, but that the articulatory output process can also benefit from previous language habits. Support for this view came from a study in which immediate memory for words and non-words was measured using either recall or recognition. Sequences of words show a clear advantage over non-words for spoken recall, but there is virtually no difference for serial-order recognition74,77. This indi- cates that language habits influence performance through the articulatory output component of the phonological loop, whereas the storage component is comparatively language independent. This could be advantageous for a system that has evolved to acquire new words, without being excessively influenced by existing knowledge. So, the phonological loop should facilitate language acquisition in two ways: the store should provide relatively unconstrained temporary representation for new phoneme sequences, and the articulatory system should facilitate learning through rehearsal, provided that the new sounds can be repre- sented using existing output processes. This faciliation is likely to occur immediately for phonotactically regular sequences, but might require further exposure when the sequences are irregular and unfamiliar.

The simple phonological loop model has proved to be robust and productive. Future developments seem likely to link it more directly to theories of language perception and production.

The visuospatial sketchpad Like its verbal equivalent, visual working memory is limited in capacity, typically to about three or four objects. This results in the phenomenon of change blindness, whereby objects in scenes can change colour, move or disappear without people noticing78,79. The visual world typically persists over time, and itself pro- vides a continuing memory record, making detailed visual retention largely redundant80. Objects comprise features such as colour, location and shape, with features within a given dimension competing for storage capac- ity, whereas features from different dimensions do not81,82. Wheeler and Triesman83 propose a model whereby feature values are stored in parallel dimension- specific registers or caches, with competition within but not between such registers. Retention of objects is dependent on the binding together of constituent features, a process that demands attention.

The visual–spatial distinction. Neuropsychological studies have indicated the need to distinguish between visual and spatial memory. The Corsi block task measures spatial span. It comprises an array of nine blocks; the experimenter initially taps two of them, and the subject attempts to imitate the sequence, with sequence length increasing until performance breaks down84 (FIG. 3a). The visual, non-spatial counterpart of this task is pattern span, in which the subject is shown matrices of cells of

a

b

c

60

64

68

72

76

80

84

88

Visual

Memory task

P er

fo rm

an ce

(% c

o rr

ec t)

Spatial

Visual interference

Spatial interference

1 2

5 6

3

7

8

4

9

Figure 3 | Visual and spatial short-term memory tasks. a | Spatial short-term memory is measured using the Corsi block tapping test. The tester taps a sequence of blocks, which the subject attempts to imitate. Sequence length increases until performance breaks down. Testing is facilitated by the numbers on the tester’s side of the blocks. Digit span is typically about two more than Corsi span. b | Spatial span involves presenting the subject with a series of matrices in which half the cells are filled. The pattern is removed and the subject is required to mark the filled cells on a response sheet. The size of the matrices is steadily increased until performance breaks down. c | Effects of interference from visual or spatial stimuli on Corsi and pattern span. Corsi span is disrupted by spatial activity and pattern span by visual activity. Panels b and c modified, with permission, from REF. 85  (1999) Elsevier Science.

834 | O C TO B E R 2 0 0 3 | VO LU M E 4 www.nature.com/reviews/neuro

R E V I E W S

whereas in the second the crucial adjectives are non-spa- tial and rely on rote verbal memory. Typically subjects can remember eight spatial sentences, but only six non- spatial ones91. The requirement to perform a concurrent tracking task, such as steering a car, disrupts the spatial code, but not the verbal condition. Although this task is predominantly spatial in character, other imagery mnemonics seem to depend much more on visual or pattern memory. Subjects who are required to learn to associate pairs of items (elephant/umbrella) can enhance performance by imagining the two interacting, so that when they are given the word elephant, the umbrella springs to mind. This kind of imagery can be disrupted by simultaneously presenting pictures of objects, patches of colour91 or even flickering patterns of visual noise92.

Function of the visuospatial sketchpad. The capacity to hold and manipulate visuospatial representations pro- vides a measure of non-verbal intelligence that predicts success in fields such as architecture and engineering93,94. There are many examples of the importance of visual or spatial imagery in scientific discovery, including Einstein’s development of his general theory of relativ- ity95. Attempts to study the design phase of engineering architecture have also identified the role of visuospatial imagery and mental synthesis93,94. In recent years, attempts to study the process of mental synthesis experi- mentally have increased. A typical study might involve presenting the subject with an array of verbally described shapes, for example, the capital letters J and D, and asking the subject to combine them into an object, with size being flexible. A suitable answer might be an umbrella96,97. The memory load can be increased by adding further shapes, and by requiring the subject to remember and manipulate them mentally as opposed to allowing the shapes to be drawn98,99. A number of studies have applied the working memory framework to such tasks, and have found that subjects will use coding in the phonological loop to store items while manipulating them visuospatially100, although this might have the disadvantage of losing visuospatial information in the process101. By analogy with the role of the phonological loop in language acquisition, it seems plausible to assume that the sketchpad might have a role in acquiring semantic knowledge about the appearance of objects and how to use them, and for understanding complex systems such as machinery, as well as for spatial orienta- tion and geographical knowledge. So far, there seems to have been little work on this potentially important topic.

Fractionation. Logie102 has proposed a fractionation of the sketchpad that is analogous to that of the loop. He distinguishes between a visual storage component, the visual cache, and a more dynamic retrieval and rehearsal process which he terms the inner scribe. He argues that the sketchpad is not a perceptually-based store, but occurs after visual information has been processed in LTM. Support for this view comes from two type

Collepals.com Plagiarism Free Papers

Are you looking for custom essay writing service or even dissertation writing services? Just request for our write my paper service, and we'll match you with the best essay writer in your subject! With an exceptional team of professional academic experts in a wide range of subjects, we can guarantee you an unrivaled quality of custom-written papers.

Get ZERO PLAGIARISM, HUMAN WRITTEN ESSAYS

Why Hire Collepals.com writers to do your paper?

Quality- We are experienced and have access to ample research materials.

We write plagiarism Free Content

Confidential- We never share or sell your personal information to third parties.

Support-Chat with us today! We are always waiting to answer all your questions.