Activation, binding, and selective access

Figure 20.1 Graphical representation of the three-layer model. The small circles represent representational elements in long-term memory, linked by associations (continuous lines); the activated part of long-term memory is designated by the filled circles. A subset of the activated elements is bound to a cognitive coordinate system, sketched as a rectangular frame; the bindings are represented by broken lines. This subset of elements constitutes the direct-access region. One of the items in the direct-access region is held in the focus of attention, denoted by the thick oval. From Oberauer (2006). Is the focus of attention in working memory expanded through practice? Journal of Experimental Psychology: Learning, Memory and Cognition, 32, 197–214. © 2006 American Psychological Association, reprinted with permission.

Retrieving a representation means taking it into the second component of working memory, the region of direct access. It constitutes the second level of selection. The direct-access region consists of a small number of the activated representations which are temporarily bound to positions in a cognitive coordinate system. A cognitive coordinate system is a context representation to which content representations can be bound. The context can be a representation of space or time, such that content elements (e.g., words, letters, objects, events, faces) can be bound to positions in the represented space, or to positions on a time axis (e.g., Burgess and Hitch 1999). A coordinate system can also be a hierarchical schema that specifies roles to be filled by content elements, for instance syntactical schemas with roles for words and phrases. Elements in the region of direct access can be accessed directly through their bindings to positions or roles in a coordinate system. For instance, the cognitive system can access the ones in a multi-digit computation by focusing on the right-most positions in mental space.

Accessing an element means bringing it into the focus of attention. This constitutes the third level of selection. The focus of attention holds the representations needed as input for the upcoming cognitive operation. It is necessary to ensure selectivity of processing in tasks that require holding several similar elements in the direct-access region, and using only a subset of them for each computational step, as in the example of multi-digit computation. The focus accesses elements in the direct-access region through their positions in the cognitive coordinate system. The contents of the direct-access region therefore form the set of candidates for selection by the focus. If representations in activated LTM need to be focused on, they must first be retrieved into the direct-access region.

Working memory is a capacity-limited system: the amount of information we can hold in it at any time is limited. In my framework there are several reasons for this limitation. One is retrieval competition between activated representations in LTM, as known from classical interference theory. There is no limit to the amount of activation to be distributed among representations in LTM -in this regard my framework differs from models such as CAPS (Just and Carpenter 1992) and ACT-R (Anderson, Reder and Lebiere 1996) – but activating too many representations can be as detrimental to efficiently retrieving the relevant information as activating too few of them, because (p.353) activated representations compete for retrieval. Inhibiting representations in LTM can therefore be advantageous (Anderson and Spellman 1995), and one source of individual differences in tasks with high working memory involvement could be the efficiency of inhibiting irrelevant information in LTM (Hasher, Zacks and May 1999; De Beni, Palladino, Pazzaglia et al. 1998). The ability to selectively inhibit activated but irrelevant representations in LTM is most relevant to tasks that require retrieval from LTM, such as across-sentence integration in reading.

A second factor relevant for the capacity of working memory is a limit to the number of independent bindings that can be held in the direct-access region. It limits the complexity of structural representations that can be set up in working memory and therefore is most relevant to reasoning tasks that require complex representations (Oberauer et al. 2007). A third limiting factor arises from the competition between elements in the direct-access region for entering the focus of attention. With more candidates in the direct-access region, selective access is slowed down. This limitation is most relevant to tasks that require quick and flexible access to individual elements in working memory for cognitive processing, such as speeded short-term retrieval tasks.

In this chapter I will summarize evidence from four sets of studies to support the framework. The first set demonstrates the usefulness of the distinction between activated LTM and the direct-access region. The second set provides evidence for all three components simultaneously. The third set of experiments shows that people can flexibly move information into and out of the direct-access region depending on whether they need to access that information. Finally, the fourth set of studies serves to link the focus of attention in working memory to the concept of a processing bottleneck (Pashler 1994).

Activation and binding in short-term recognition

Activation of a representation in LTM and binding of a representation to a context in the direct-access region play different roles in short-term recognition tasks. These roles can be mapped to the two processes assumed in dual-process theories of recognition (Atkinson, Herrmann and Wescourt 1974; Yonelinas 2002). One process, assumed to be fast and automatic, is the assessment of the familiarity of a probe (i.e., a stimulus to be judged as old or new). The other process, assumed to be slower and controlled, consists of recollecting information about previous encounters with the probe. I assume that familiarity assessment is based on a measurement of how quickly and fluently the probe is processed. In short-term recognition tasks, where only a few seconds separate encoding and probing of a stimulus, processing fluency is largely determined by the current level of activation of a representation of the probe in LTM. In recognition over longer time intervals familiarity is most likely driven by the strength of associations between features of the probe, so-called intra-item associations (Mandler 1980). Recollection involves retrieving that the probe has been encountered before in the relevant context – for instance, retrieving that a word has been on the last list studied (as opposed to a previous list, or in a book read before the experiment). Recollection therefore rests on retrieving a link between the content of the probe stimulus and the context of its previous encounter (if there was one), or on retrieving that there was no such element in the relevant context (if there was none). In short-term recognition tasks, I assume that this information is mostly provided by the temporary bindings between content elements and context representations in the direct-access region.

Familiarity and recollection can be separated by conflict recognition tasks (e.g., Jennings and Jacoby 1997). These are paradigms in which, for a subset of probes, familiarity and recollection support conflicting recognition decisions. An example is the modified Sternberg task (Oberauer 2001), illustrated in Figure 20.2. Participants try to remember two short lists of words, after which they receive a cue telling them which of the two lists is relevant for the upcoming recognition decision.(p.354)

Activation, binding, and selective accessAn embedded three-component framework for working memory

Figure 20.2 Sequence of events in a trial of the modified Sternberg task (Oberauer 2001). Two lists of words are presented simultaneously as two lines, one in blue and one in red (illustrated here through different fonts). This is followed by a frame colored red or blue, indicating which list is relevant for this trial. After a variable cue-stimulus interval (CSI), the probe appears, and reaction time is measured until the participant decides through a key-press whether the probe was in the relevant list. The example shows an intrusion trial.

After a variable cue-stimulus interval (CSI), a probe appears and participants decide whether it had been in the relevant list. Probes that had been in the irrelevant list (but not in the relevant list) are assessed as highly familiar because, up to the time where the cue is displayed, their representations are activated in LTM just as much as those of the relevant list. Recollection, however, can provide information about which list the probe was in, which eventually leads to correct rejection of these so-called intrusion probes. This requires that at least the relevant list is held in the region of direct access.

In an experiment involving young and old adults I investigated the time course of removing the irrelevant list from working memory (Oberauer 2001). The CSI was varied in six steps from 100 ms to 5 s. With increasing CSI, participants have more time to remove the irrelevant list from the direct-access region, thereby taking this load off the limited capacity to maintain bindings, and reducing the competition for access to elements in the relevant list. In addition, the activation of irrelevant-list items in LTM can decline over the CSI, either through passive decay or active inhibition. It turned out that both young and old adults were very efficient in removing the irrelevant list from the direct-access region. After about 1 s after the cue, there was no effect of list-length of the irrelevant list on reaction times to the probe. This indicates that elements from the irrelevant list no longer burdened the limited capacity of the direct-access region. There was no significant age difference in the speed of reducing the list-length effect of the irrelevant list to zero in that study, but later investigations with the same paradigm suggested that old adults might be somewhat less effective in this regard (Oberauer 2005b). Reducing the activation of irrelevant-list representations in LTM, in contrast, was much slower. Even after 5 s, participants had considerably more difficulties with rejecting intrusion probes than with rejecting new probes that were not part of any list, demonstrating a persistent conflict between familiarity and recollection. Moreover, these difficulties were disproportionally increased in old compared to young adults. These results are displayed in Figure 20.3.

(p.355)

Figure 20.3 Top panel: Irrelevant-set size effect in the modified Sternberg task. Data points reflect differences between RTs on trials with irrelevant-set size 3 and trials with irrelevant-set size 1. Bottom panel: Intrusion costs in the modified Sternberg task. Data points represent differences between RTs on intrusion trials and RTs on negative trials (Oberauer 2001). Error bars reflect two standard errors (i.e., approximately 95 per cent confidence intervals).

Age differences in the difficulty of rejecting intrusion probes could arise from two sources. Old adults could have more difficulties with inhibiting irrelevant representations in LTM (Hasher et al. 1999). Alternatively, old adults could have problems with maintaining robust content–context bindings in the direct-access region. This deficit would compromise their recollection, leaving the misleading familiarity of intrusion probes more time to build up and influence recognition decisions. (p.356) A recent follow-up study aimed at distinguishing these possibilities (Oberauer, 2005b). I tested young and old adults with three conflict recognition tasks, the modified Sternberg task, an n-back task (Kirchner 1958), and a local recognition task. The n-back task required participants to compare each word in a 20-word sequence to the word n steps back, withn varying between one and three. The local recognition task required them to remember a list of items, each item being displayed in a separate frame. Probes were displayed in the frames and had to be compared to the list items in these frames (see Figure 20.4). In both tasks, intrusion probes were included. In the n-back task, intrusion probes were words that matched a previous word less than n steps back. These words can be assumed to be highly activated and therefore highly familiar, but they must be rejected on the basis of recollecting their ordinal position on a temporal context vector (e.g., remembering that the present word had occurred one step before instead of two). In the local recognition task, intrusion probes were probes that had been in the memory list but in a different frame. Again, these probes match a highly activated representation but must be rejected through recollection, based on the bindings between list items and their frames (i.e., their positions in a spatial coordinate system).

An important feature of the n-back task and the local recognition task that sets them apart from the modified Sternberg task is that the representations generating the familiarity signal on intrusion probes must not be inhibited, because they are not permanently irrelevant. Words less than n steps back become relevant once they are n steps back. Words that had been presented in a different frame from the one currently probed in the local recognition task will become relevant when that

Figure 20.4 Sequence of events in a trial of the local recognition task with set size 3. Three letters are presented sequentially, each in one frame, for 1000 ms each. After a 500 ms blank, three probes appear in red (represented by bold letters). In the example, the first probe is a positive probe, the second an intrusion probe, and the third is a negative probe.

(p.357) frame is probed later. If old adults are at a disadvantage in the modified Sternberg task because of an inhibition deficit, they should not be particularly impaired relative to young adults on these other conflict recognition task, because young adults could not benefit on these tasks from their more efficient inhibition mechanisms. If old adults’ problems with the modified Sternberg task arise from deficits in recollection, in contrast, they should show the same problems with the other two conflict recognition tasks, too. The latter was the case: in all three conflict recognition tasks, old adults’ reaction times on intrusion probes were disproportionally slowed compared to young adults’. The impairment of old adults’ performance on conflict recognition tasks therefore seems to arise from a reduced capacity of the direct-access region to maintain robust content–context bindings, not from a deficit in actively reducing the activation of irrelevant representations in LTM.

I further investigated individual differences in familiarity assessment and recollection in the young group (Oberauer, 2005b). A structural-equation model analysis showed that performance in conflict recognition tasks can be decomposed into two sources of variance, one reflecting familiarity and the other reflecting recollection. The recollection factor but not the familiarity factor was correlated with a latent factor representing working memory capacity, as assessed by four established measures of that construct. This finding supports the hypothesis that working memory capacity, as reflected in performance on tasks such as reading span (Daneman and Carpenter 1980), is essentially the capacity to hold bindings in the region of direct access (Oberauer et al. 2007).

Observing all three components in action

My main empirical reason for distinguishing between the direct-access region and the activated part of LTM arises from findings concerning list-length effects on reaction times. An increase of reaction times with the length of the list to be held in working memory is a common finding, the most famous example being the linear slope of RTs over list length in the Sternberg (1969) task that I exploited in the first study reviewed here. List length, however, does not always affect RTs. Results from several studies support the generalization that RTs increase with list length if and only if the RT task requires access to information from the list in working memory (Kessler and Meiran 2006; Oberauer, Demmrich, Mayr et al. 2001; Oberauer 2002, 2005a). This can be illustrated by an experiment using an arithmetic task (Oberauer 2002, Experiment 2). In each trial participants had to remember two short lists of digits; each list could consist of one or of three digits. Each digit was presented in a separate frame. Following encoding, a cue indicated which of the two lists would be the ‘active’ list. After a variable CSI, a series of arithmetic operations had to be completed, using digits from the active list as input. Stimuli for these processing tasks were operation instructions such as ‘+ 2’ or ‘− 6’ displayed in individual frames of the active list. Participants had to retrieve the digit that had been presented in the appropriate frame, apply the operation to it, and type the result as quickly as possible. During the whole series of eight arithmetic operations they had to remember not only the digits in the active list but also the digits in the other, passive list, because at the end of the trial they had to recall the digits of both lists (see Figure 20.5). Participants were able to recall both lists reasonably well, so both lists must be assumed to be held in working memory (on successful trials). The important finding was that the length of the active list always had an effect on the latency of individual arithmetic operations. The length of the passive list, however, only affected the first operation in the series, and only on short CSIs. In other words, after about three seconds following the cue, there was no list-length effect of the passive list on arithmetic processing times (see Figure 20.6).

The three-component framework offers a straightforward explanation for this observation: because the processing task requires access to memory elements on the active list, that the list (p.358)

Figure 20.5 Sequence of events in a trial of the arithmetic task used in Oberauer (2002, Experiment 2). After initial encoding of two lists – here both with set size 3 – a cue indicated which list would be the active one (red frames, here marked by thick lines). Eight arithmetic operations were performed on digits of the active list. The second operation in the example represents an object switch. At the end, the digits in each frame are probed for recall.

Figure 20.6 Reaction times in the arithmetic task of Oberauer (2002, Experiment 2) by set size of the passive set, for the first operation in the sequence. Error bars represent 95 per cent confidence intervals for within-subject comparisons (Bakeman and McArthur 1996).

Only gold members can continue reading. Log In or Register to continue