Working Memory Model Research Articles

A biophysical neural network model for visual working memory that accounts for memory binding errors

Binding errors, also called swap errors, occur in working memory tasks when the participant fails to report the feature of a previously presented target but the response is accurate relative to a non-target stimulus (for instance, s/he reports the of the green item when the red item's was to be reported). Psychophysical experiments have shown the importance of these errors in working memory tasks [1-3], and they reflect the failure of the system to maintain bundled through memory the conjunction of features that define one object. How the brain is able to integrate and maintain integrated several features of an item in one single memory is still an open question. Conjunctive selectivity, where single units are selective to combination of stimuli, has been suggested as a possible solution [4]. However, with the increase of features to encode the number of necessary neurons escalates exponentially. A hypothesis formulated for this problem in visual perception states that synchrony of different neural assemblies coding each of them a single feature of an item plays the main role [5]. This is an attractive hypothesis, because of its parsimonious and flexible neural representation: a particular neuron can participate in the coding of different objects at different times. Because under this hypothesis there is no hardwired selectivity for each group of features, it allows ex novo representations to emerge. Current neurophysiological data provides partial support for this idea but the specific neural mechanism remains unknown [5,6]. In order to understand this phenomenon and to test this synchrony hypothesis, we built a biophysical neural network model for the storage of items orks for working memory (ring models, as in [7]), one representing colors and the other one locations of equal eccentricity. These two networks are then connected via weak cortico-cortical excitation through AMPA synapses. With this model we are able to maintain persistent activity through recurrent synaptic interactions in memory [8] that encode colors and locations. AMPA synapses in recurrent excitation within each network induce gamma oscillations during bump activity through the interplay of fast excitation and slower feedback inhibition [7]. We found that when multiple are maintained within one ring model, the corresponding bump attractors oscillate out of synchrony with each other. In this model, the binding between and is effectively accomplished through the synchronization of gamma oscillations between individual in the two networks as a result of weak cortico-cortical excitation. As previously proposed [9-10], in our network model different memories are held at different phases of a low-frequency oscillation. In some simulations swap errors arise: color bumps abruptly change their phase and thus change their synchrony relationship with location bumps. This leads to a wrong association between and that may underlie memory binding errors reported in working memory tasks.

The neurodynamical basis of multi-item working memory capacity: sequential vs simultaneous stimulation paradigms

When investigating multi-item WM, and in contrast to single item experiments, a decision must be made regarding a key aspect of the stimulation protocol: how the memory set is presented to the subject simultaneously or sequentially. It is worth noting that most studies investigating multi-item WM do not address this issue and focus either in simultaneous stimulation protocols (e.g. [1,2]) or in sequential stimulation protocols (e.g. [3]) without confronting the two situations. This is nevertheless an aspect which provides a benchmark to probe and compare the different theories regarding how resources are allocated among the different items of a memory set [4,5]. In this study, we explore a biophysically-realistic attractor model of visual working memory (VWM) endowed with synaptic facilitation and investigate what are the effects of varying the dynamics of the facilitation process. We fi nd that: 1) it is possible to reproduce experimentally observed effects such as the recency effect in sequential stimulation protocols (i.e. items presented in the final positions of a sequence are more likely to be retained in WM), and 2) WM capacity is boosted in both sequential and stimulation protocols when endowing the attractor network with synaptic facilitation. Conclusions In agreement with our previous results [2], synaptic facilitation boosts the WM capacity limit by effectively increasing the synaptic strengths just for those pools to which a cue is applied, and then maintaining the synaptic facilitation by the continuing neuronal firing in only these pools when the cue is removed. In this study, the time constant τF of the synaptic facilitation process has been found to play an important role in modulating this effect with large τF values leading to larger capacity limits in both sequential and simultaneous stimulation protocols. However, too large τF values lead to neuronal dynamics which are not compatible with the recency effect, thus constraining the range of values that τF may take.

Double dissociation of a latent working memory process

ABSTRACTBackground. The study explored the construct validity of a computational model of working memory (WM) by determining whether model parameters manifested double dissociations of lesion laterality with type of material studied. The data set modeled involved psychometrically matched verbal and figural WM tasks on which a double dissociation between test version and lesion laterality failed to emerge when total test scores were used as the laterality marker. Method. This re-analysis of a previously published study involved investigating the WM performance of 15 demographically matched controls with 15 adult patients with left-hemispheric (LH) lesions and 15 adult patients with right-hemispheric (RH) lesions. Each participant was given verbal and figural versions of a continuous paired associates test (CPAT). The two versions had previously been psychometrically matched in a larger sample of healthy individuals. A WM model composed of encoding, displacement, and episodic memory parameters was fit to each individual’s performance profiles for both versions of the CPAT. Results. Replicating the previous results for raw scores, rank transformed values of total score performance failed to reveal a double dissociation. Nonetheless an absolute double dissociation was observed for the model’s displacement parameter: RH patients demonstrated deficits on the figural but not verbal WM displacement parameter, whereas LH patients demonstrated deficits on the verbal but not figural WM displacement parameter. Additionally, both LH and RH patients were impaired on the figural encode parameter, perhaps explaining the absence of a double dissociation in total score performance. Conclusions. By combining different patterns of profile and level data into theoretically motivated model parameters, computational models of neurocognition can enhance the construct validity of interpretations of neuropsychological performance.

Contextual effects in visual working memory reveal hierarchically structured memory representations.

Influential slot and resource models of visual working memory make the assumption that items are stored in memory as independent units, and that there are no interactions between them. Consequently, these models predict that the number of items to be remembered (the set size) is the primary determinant of working memory performance, and therefore these models quantify memory capacity in terms of the number and quality of individual items that can be stored. Here we demonstrate that there is substantial variance in display difficulty within a single set size, suggesting that limits based on the number of individual items alone cannot explain working memory storage. We asked hundreds of participants to remember the same sets of displays, and discovered that participants were highly consistent in terms of which items and displays were hardest or easiest to remember. Although a simple grouping or chunking strategy could not explain this individual-display variability, a model with multiple, interacting levels of representation could explain some of the display-by-display differences. Specifically, a model that includes a hierarchical representation of items plus the mean and variance of sets of the colors on the display successfully accounts for some of the variability across displays. We conclude that working memory representations are composed only in part of individual, independent object representations, and that a major factor in how many items are remembered on a particular display is interitem representations such as perceptual grouping, ensemble, and texture representations.

Dual-task backward compatibility effects are episodically mediated.

In dual-task performance, the backward compatibility effect (BCE; faster Task 1 reaction time when Task 1 and Task 2 responses are compatible) is thought to represent automatic activation of Task 2 response information in parallel with attended Task 1 performance. Work by Hommel and Eglau (Psychological Research, 66, 260-273, 2002) has suggested the BCE relies on stimulus-response learning in long-term memory. Subsequent work by Ellenbogen and Meiran (Memory and Cognition, 36, 968-978, 2008), however, proposed that the BCE is mediated by Task 2 rules held in working memory (WM) during Task 1 performance. The present study aimed to dissociate these two theoretical claims. In Experiment 1, we assessed the effects of prior single-task practice with Task 1 or Task 2 of a subsequent dual-task paradigm. Where the WM-mediated model predicts both BCE and overall reaction time improvement relative to prior task practice, an episodic learning model makes divergent predictions for BCE based on the context specificity of prior Task 2 learning. Results showed a close fit with episodic predictions and contradicted WM model predictions. Experiment 2 examined the finer grained timecourse of BCE over initial development, subsequent interference of this initial learning on BCE development with new conflicting Task 2 response mappings, and finally reestablishment of BCE in the original dual task. Data again showed close agreement with long-term learning predictions. We argue in favor of an episodic account of the BCE, and consider implications of WM and episodic mechanisms of automatic response activation on other aspects of dual-task performance.

Editorial: Turning the Mind's Eye Inward: The Interplay Between Selective Attention and Working Memory.

Editorial: Turning the Mind's Eye Inward: The Interplay Between Selective Attention and Working Memory.

Neurocognitive predictors of mathematical processing in school-aged children with spina bifida and their typically developing peers: Attention, working memory, and fine motor skills.

Math and attention are related in neurobiological and behavioral models of mathematical cognition. This study employed model-driven assessments of attention and math in children with spina bifida myelomeningocele (SBM), who have known math difficulties and specific attentional deficits, to more directly examine putative relations between attention and mathematical processing. The relation of other domain general abilities and math was also investigated. Participants were 9.5-year-old children with SBM (n = 44) and typically developing children (n = 50). Participants were administered experimental exact and approximate arithmetic tasks, and standardized measures of math fluency and calculation. Cognitive measures included the Attention Network Test (ANT), and standardized measures of fine motor skills, verbal working memory (WM), and visual-spatial WM. Children with SBM performed similarly to peers on exact arithmetic, but more poorly on approximate and standardized arithmetic measures. On the ANT, children with SBM differed from controls on orienting attention, but not on alerting and executive attention. Multiple mediation models showed that fine motor skills and verbal WM mediated the relation of group to approximate arithmetic; fine motor skills and visual-spatial WM mediated the relation of group to math fluency; and verbal and visual-spatial WM mediated the relation of group to math calculation. Attention was not a significant mediator of the effects of group for any aspect of math in this study. Results are discussed with reference to models of attention, WM, and mathematical cognition.

Examining the role of working memory resources in following spoken instructions

ABSTRACTThis study investigated the involvement of working memory (WM) in following spoken instructions using concurrent tasks designed to disrupt components of the Baddeley and Hitch WM model [Baddeley, A. D., & Hitch, G. J. (1974). Working memory. In G. H. Bower (Ed.), The psychology of learning and motivation: Advances in research and theory (Vol. 8, pp. 47–89). New York, NY: Academic Press]. In each of three experiments, participants were presented with sequences of instructions to be either verbally repeated or physically performed using relevant objects. Backward counting, articulatory suppression, and eye closure during instruction encoding all disrupted recall, and also impaired recall of the linkage between movements and objects. Recall of actions was more accurate when they were physically enacted than repeated verbally, an advantage that was not affected by concurrent tasks. These findings indicate that aspects of the recall of spoken instructions including the binding of constituent movements to objects draw on multiple WM resources. The benefits of physical enactment of the instructed sequence do not appear to depend on the components of WM investigated in these studies.

Attention, working memory, and phenomenal experience of WM content: memory levels determined by different types of top-down modulation.

What is the role of top-down attentional modulation in consciously accessing working memory (WM) content? In influential WM models, information can exist in different states, determined by allocation of attention; placing the original memory representation in the center of focused attention gives rise to conscious access. Here we discuss various lines of evidence indicating that such attentional modulation is not sufficient for memory content to be phenomenally experienced. We propose that, in addition to attentional modulation of the memory representation, another type of top-down modulation is required: suppression of all incoming visual information, via inhibition of early visual cortex. In this view, there are three distinct memory levels, as a function of the top-down control associated with them: (1) Nonattended, nonconscious associated with no attentional modulation; (2) attended, phenomenally nonconscious memory, associated with attentional enhancement of the actual memory trace; (3) attended, phenomenally conscious memory content, associated with enhancement of the memory trace and top-down suppression of all incoming visual input.

Efficient reinforcement learning of a reservoir network model of parametric working memory achieved with a cluster population winner-take-all readout mechanism

The brain often has to make decisions based on information stored in working memory, but the neural circuitry underlying working memory is not fully understood. Many theoretical efforts have been focused on modeling the persistent delay period activity in the prefrontal areas that is believed to represent working memory. Recent experiments reveal that the delay period activity in the prefrontal cortex is neither static nor homogeneous as previously assumed. Models based on reservoir networks have been proposed to model such a dynamical activity pattern. The connections between neurons within a reservoir are random and do not require explicit tuning. Information storage does not depend on the stable states of the network. However, it is not clear how the encoded information can be retrieved for decision making with a biologically realistic algorithm. We therefore built a reservoir-based neural network to model the neuronal responses of the prefrontal cortex in a somatosensory delayed discrimination task. We first illustrate that the neurons in the reservoir exhibit a heterogeneous and dynamical delay period activity observed in previous experiments. Then we show that a cluster population circuit decodes the information from the reservoir with a winner-take-all mechanism and contributes to the decision making. Finally, we show that the model achieves a good performance rapidly by shaping only the readout with reinforcement learning. Our model reproduces important features of previous behavior and neurophysiology data. We illustrate for the first time how task-specific information stored in a reservoir network can be retrieved with a biologically plausible reinforcement learning training scheme.

Precision of working memory for speech sounds.

Memory for speech sounds is a key component of models of verbal working memory (WM). But how good is verbal WM? Most investigations assess this using binary report measures to derive a fixed number of items that can be stored. However, recent findings in visual WM have challenged such "quantized" views by employing measures of recall precision with an analogue response scale. WM for speech sounds might rely on both continuous and categorical storage mechanisms. Using a novel speech matching paradigm, we measured WM recall precision for phonemes. Vowel qualities were sampled from a formant space continuum. A probe vowel had to be adjusted to match the vowel quality of a target on a continuous, analogue response scale. Crucially, this provided an index of the variability of a memory representation around its true value and thus allowed us to estimate how memories were distorted from the original sounds. Memory load affected the quality of speech sound recall in two ways. First, there was a gradual decline in recall precision with increasing number of items, consistent with the view that WM representations of speech sounds become noisier with an increase in the number of items held in memory, just as for vision. Based on multidimensional scaling (MDS), the level of noise appeared to be reflected in distortions of the formant space. Second, as memory load increased, there was evidence of greater clustering of participants' responses around particular vowels. A mixture model captured both continuous and categorical responses, demonstrating a shift from continuous to categorical memory with increasing WM load. This suggests that direct acoustic storage can be used for single items, but when more items must be stored, categorical representations must be used.

Temporal dynamics of visual working memory

The involvement of the human cerebellum in working memory has been well established in the last decade. However, the cerebro-cerebellar network for visual working memory is not as well defined. Our previous fMRI study showed superior and inferior cerebellar activations during a block design visual working memory task, but specific cerebellar contributions to cognitive processes in encoding, maintenance and retrieval have not yet been established. The current study examined cerebellar contributions to each of the components of visual working memory and presence of cerebellar hemispheric laterality was investigated. 40 young adults performed a Sternberg visual working memory task during fMRI scanning using a parametric paradigm. The contrast between high and low memory load during each phase was examined. We found that the most prominent activation was observed in vermal lobule VIIIb and bilateral lobule VI during encoding. Using a quantitative laterality index, we found that left-lateralized activation of lobule VIIIa was present in the encoding phase. In the maintenance phase, there was bilateral lobule VI and right-lateralized lobule VIIb activity. Changes in activation in right lobule VIIIa were present during the retrieval phase. The current results provide evidence that superior and inferior cerebellum contributes to visual working memory, with a tendency for left-lateralized activations in the inferior cerebellum during encoding and right-lateralized lobule VIIb activations during maintenance. The results of the study are in agreement with Baddeley's multi-component working memory model, but also suggest that stored visual representations are additionally supported by maintenance mechanisms that may employ verbal coding.

Using Bayes factors to test the predictions of models: A case study in visual working memory

A critical property of Bayesian model selection, via Bayes factors, is that they test the predictions made by models. Such predictions are a joint function of the likelihood of the model, and the prior distributions placed on the parameters of the model. Prior distributions that are informed by previous data lead to more constrained predictions, and result in Bayes factors that test more specific versions of the models under question. We present a case study applying two models of visual working memory to a series of experiments. We outline a process by which the posterior distributions from previous experiments are used to define and update prior distributions for each subsequent experiment. For each experiment we obtain Bayes factors that test the predictions that these models make, and update our beliefs about the relative likelihood of each model.

Parietal and Frontal Cortex Encode Stimulus-Specific Mnemonic Representations during Visual Working Memory

Working memory (WM) enables the storage and manipulation of information in an active state. WM storage has long been associated with sustained increases in activation across a network of frontal and parietal cortical regions. However, recent evidence suggests that these regions primarily encode information related to general task goals rather than feature-selective representations of specific memoranda. These goal-related representations are thought to provide top-down feedback that coordinates the representation of fine-grained details in early sensory areas. Here, we test this model using fMRI-based reconstructions of remembered visual details from region-level activation patterns. We could reconstruct high-fidelity representations of a remembered orientation based on activation patterns in occipital visual cortex and in several sub-regions of frontal and parietal cortex, independent of sustained increases in mean activation. These results challenge models of WM that postulate disjoint frontoparietal "top-down control" and posterior sensory "feature storage" networks.

Neural circuit basis of visuo-spatial working memory precision: a computational and behavioral study.

The amount of information that can be retained in working memory (WM) is limited. Limitations of WM capacity have been the subject of intense research, especially in trying to specify algorithmic models for WM. Comparatively, neural circuit perspectives have barely been used to test WM limitations in behavioral experiments. Here we used a neuronal microcircuit model for visuo-spatial WM (vsWM) to investigate memory of several items. The model assumes that there is a topographic organization of the circuit responsible for spatial memory retention. This assumption leads to specific predictions, which we tested in behavioral experiments. According to the model, nearby locations should be recalled with a bias, as if the two memory traces showed attraction or repulsion during the delay period depending on distance. Another prediction is that the previously reported loss of memory precision for an increasing number of memory items (memory load) should vanish when the distances between items are controlled for. Both predictions were confirmed experimentally. Taken together, our findings provide support for a topographic neural circuit organization of vsWM, they suggest that interference between similar memories underlies some WM limitations, and they put forward a circuit-based explanation that reconciles previous conflicting results on the dependence of WM precision with load.

Working memory impairment and its associated sleep-related respiratory parameters in children with obstructive sleep apnea

Study ObjectiveWorking memory deficits in children with obstructive sleep apnea (OSA) have been reported in previous studies, but the results were inconclusive. This study tried to address this issue by delineating working memory functions into executive processes and storage/maintenance components based on Baddeley's working memory model. MethodsWorking memory and basic attention tasks were administered on 23 OSA children aged 8–12 years and 22 age-, education-, and general cognitive functioning-matched controls. Data on overnight polysomnographic sleep study and working memory functions were compared between the two groups. Associations between respiratory-related parameters and cognitive performance were explored in the OSA group. ResultsCompared with controls, children with OSA had poorer performance on both tasks of basic storage and central executive components in the verbal domain of working memory, above and beyond basic attention and processing speed impairments; such differences were not significant in the visuo-spatial domain. Moreover, correlational analyses and hierarchical regression analyses further suggested that obstructive apnea–hypopnea index (OAHI) and oxygen saturation (SpO2) nadir were associated with verbal working memory performance, highlighting the potential pathophysiological mechanisms of OSA-induced cognitive deficits. ConclusionsVerbal working memory impairments associated with OSA may compromise children's learning potentials and neurocognitive development. Early identification of OSA and assessment of the associated neurocognitive deficits are of paramount importance. Reversibility of cognitive deficits after treatment would be a critical outcome indicator.

A comparative study of the working memory multicomponent model in psychosis and healthy controls

Working memory deficits are considered nuclear deficits in psychotic disorders. However, research has not found a generalized impairment in all of the components of working memory. We aimed to assess the components of the Baddeley and Hitch working memory model: the temporary systems—the phonological loop, the visuospatial sketchpad and the episodic buffer (introduced later by Baddeley)—and the central executive system, which includes four executive functions: divided attention, updating, shifting and inhibition. We assessed working memory performance in a sample of 21 patients with a psychotic disorder and 21 healthy controls. Patients also underwent a clinical assessment. Both univariate and repeated measures ANOVAs were applied to analyze performance in the working memory components between groups. Patients with a psychotic disorder underperformed compared to the controls in all of the working memory tasks, but after controlling for age and premorbid IQ, we only found a difference in performance in the N-Back task. Repeated measures ANCOVAs showed that patients also underperformed compared to the controls in the Digit span test and the TMT task. Not all of the components of working memory were impaired in the patients. Specifically, patients' performance was impaired in the tasks selected to assess the phonological loop and the shifting executive function. Patients' also showed worse performance than controls in the N-Back task, representative of the updating executive function. However, we did not find higher impairment in the patients' performance respect to controls when increasing the loading of the task.

Forward scanning in verbal working memory updating.

Effective use of working memory (WM) for high-level cognitive tasks requires coordinating two conflicting requirements: robust maintenance and rapid updating. Models of WM suggest that these demands are coordinated by a gate between perceptual input and WM. Previous work with a letter-updating paradigm (Kessler & Oberauer, Journal of Experimental Psychology: Learning, Memory, and Cognition, 40, 738-754, 2014) supported a scanning and gate-switching (SGS) model of WM updating. The present work provides further evidence for the SGS model. Participants were required to keep track of the last letter that appeared in each of a row of frames on the screen. On each updating step, a variable subset of letters in varying positions in the row had to be updated. The SGS model assumes that on each updating step, participants scan through the memory set sequentially, opening the gate when a letter requires updating, and closing the gate when the next letter needs to be maintained. As is predicted by the SGS model, the reaction times for each updating step increased with the number of updated items and with the number of gate switches. In addition, the present experiment provides direct evidence supporting the scanning assumption of the model. Hebrew-speaking participants performed the task with either Hebrew or English letter stimuli, in different blocks. As was predicted, the scanning direction of the stimulus set was from left to right in English and from right to left in Hebrew. The SGS model fit the data only when the scanning direction was taken into account, establishing the role of item-based forward scanning during WM updating.

Measuring Working Memory With Digit Span and the Letter-Number Sequencing Subtests From the WAIS-IV: Too Low Manipulation Load and Risk for Underestimating Modality Effects

The Wechsler Adult Intelligence Scale (WAIS) is one of the most frequently used tests among psychologists. In the fourth edition of the test (WAIS-IV), the subtests Digit Span and Letter-Number Sequencing are expanded for better measurement of working memory (WM). However, it is not clear whether the new extended tasks contribute sufficient complexity to be sensitive measures of manipulation WM, nor do we know to what degree WM capacity differs between the visual and the auditory modality because the WAIS-IV only tests the auditory modality. Performance by a mixed sample of 226 patients referred for neuropsychological examination on the Digit Span and Letter-Number Sequencing subtests from the WAIS-IV and on Spatial Span from the Wechsler Memory Scale-Third Edition was analyzed in two confirmatory factor analyses to investigate whether a unitary WM model or divisions based on modality or level/complexity best fit the data. The modality model showed the best fit when analyzing summed scores for each task as well as scores for the longest span. The clinician is advised to apply tests with higher manipulation load and to consider testing visual span as well before drawing conclusions about impaired WM from the WAIS-IV.

The Multicomponent Working Memory Model, Attention, and Long-term Memory in Multimedia Learning: A Comment on Schweppe and Rummer (2014)

Multimedia learning theories rely on findings from different areas of fundamental psychological research. Working memory is undoubtedly a particularly important concept for these theories. However, the application of findings from fundamental research to multimedia learning is not as straightforward as it might seem at first glance, and misinterpretations are pervasive in the literature. The recent article of Schweppe and Rummer (2014) represents an important contribution in this respect. Their paper outlines assumptions in Baddeley and Hitch’s (1974) multicomponent working model and discusses the implementation of these assumptions in popular multimedia learning theories such as the cognitive theory of multimedia learning (CTML; Mayer 2009) and cognitive load theory (CLT; Sweller et al. 2011). The authors then continue arguing that the multicomponent model conceives working memory as structurally separate from long-term memory, and that this assumed separation has drawbacks with regard to explaining a number of effects in multimedia studies. As an alternative foundation for multimedia learning theories, they propose Cowan’s (1999) embedded-processes model. While this effort should be highly valued, the present commentary argues (a) that Schweppe and Rummer’s rejection of the multicomponent model as a basis for multimedia theories is based on a misinterpretation of the relationship between working memory and long-term memory in the multicomponent model, (b) that the multicomponent model and the embeddedprocesses model can be seen as complementary approaches, and (c) that a complete theory of multimedia learning needs to fully integrate both central executive (i.e., attentional) and functionally specialized (e.g., phonological and visuospatial) mechanisms.