Event Cache: An Independent Component in Working Memory

doi:10.21203/rs.3.rs-2745694/v1

Download PDF

Article

Event Cache: An Independent Component in Working Memory

https://doi.org/10.21203/rs.3.rs-2745694/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Working memory (WM) has been a major focus of cognitive science and neuroscience for the past 50 years. While most WM research has centered on the mechanisms of objects, there has been a lack of investigation into the cognitive and neural mechanisms of events, which are the building blocks of our experience. Employing confirmatory factor analysis and resting-state and task fMRI, our study demonstrated for the first time that events have an independent storage space within WM, known as the event cache, with distinct neural correlates compared to object storage in WM. The cerebellar network was found to be the most essential network for event cache, with the left cerebellum Crus I being particularly involved in encoding and maintaining events. Our findings shed critical light on the neuropsychological mechanism of WM by revealing event cache as an independent sub-component of WM and encourage the reconsideration of theoretical models for WM.

Biological sciences/Psychology/Human behaviour

Biological sciences/Neuroscience/Cognitive neuroscience/Cognitive control

The ability to reduce everything to simple fundamental laws does not imply the ability to start from those laws and reconstruct the universe. …The behavior of large and complex aggregates of elementary particles, it turns out, is not to be understood in terms of a simple extrapolation of the properties of a few particles. Instead, at each level of complexity entirely new properties appear, and the understanding of the new behaviors requires research which I think is as fundamental in its nature as any other.

P. W. Anderson in More is different, 1972

Working memory (WM) refers to a system that temporarily stores and manipulates a limited set of information for ongoing tasks. ^1–3 Although WM capacity is limited, it is essential for high-order cognitions, such as fluid intelligence, ^4,5 reading comprehension, ⁶ planning, ⁷ and creative thinking. ⁸ WM deficits have been reported in various mental disorders such as Alzheimer’s disease, attention deficit hyperactivity disorder, social anxiety disorder, and schizophrenia. ^9–12 Understanding the mechanisms of WM has significant educational and clinical implications and has been a key topic in cognitive science and neuroscience for the past 50 years. ^13–17

Different WM theories have been proposed to elucidate how information is stored and manipulated in WM. Two of the most influential models are the multi-component model and the embedded model. ^1–3 The multi-component model claims that WM consists of four components: central executive (CE), phonological loop, visuo-spatial sketchpad, and episodic buffer. ^2,18 The CE is responsible for coordinating and monitoring ongoing tasks in a top-down manner, whereas the remaining three are responsible for retaining verbal, visuo-spatial, and binding information, respectively. ¹⁸ Moreover, these distinct components are supported by different neural substrates or networks. ² In contrast, the embedded model and its variants suggest that all information is stored in a capacity-limited focus of attention (FoA), which is a subset of information in long-term memory that is temporarily activated. ^19–21

Due to the reductionism routine and past technical limitations in psychological explorations, the main body of WM studies, including theoretical models, have focused on verbal and visual objects and their constituent features. ^2,22 In one’s daily life, one not only processes objects and features, but also pays attention to what happens to them — and what happens is an event, corresponding to a psychological representation format distinct from objects. ^23,24 For example, the biological motion of humans is one of the most typical events in daily life. ^25,26 Events are fundamental units of our experience, and our experience of the world reflects the way we connect and separate various events. Although events manifest dynamically, only changes that contain higher-order stability through their duration are individuated as events. ²⁴ The brain has evolved to use temporal structures in events to predict incoming information and enable functional behaviors. Consequently, events are considered to be one of the most important classes of entities in daily psychology, ^23,24,27,28 and should be a critical unit of analysis for most psychological domains, including WM.

Indeed, recent studies have revealed that continuous information flow from the environment is segmented into events in the WM. ^29–31 Surprisingly, studies of events have predominantly explored event processing in perception and long-term memory. ^27,28 Neither the psychological construct nor the underlying neural substrates of event processing in WM are well understood. In particular, whether there is a unique cache for events in WM has not yet been investigated. As an event is situated in a higher representation node than an object in terms of representational hierarchy in the mind, and at each representation level of mind, entirely new properties appear according to the emergence principle of Anderson, ³² and an event hence may have a distinct storage space from the object.

To address whether events have a separate storage space in mind compared to objects, we selected biological movements of humans and non-biological movements of physical shapes as the representatives of events and adopted three converging approaches (Fig. 1). First, we adopted confirmatory factor analysis (CFA) to build models based on a battery of 14 well-established WM tasks to examine whether event cache could be established as a unique psychometric construct distinct from other well-known storage buffers. CFA can represent each cognitive construct at the latent factor level by extracting the common variance among multiple tasks that are selected to tap the same underlying construct, which reduces the task impurity problem in measuring cognitive constructs. Second, since resting-state fMRI is able to reliably predict individual differences in cognitive ability, such as WM capacity and attention, ^33–35 we applied a support vector regression (SVR) algorithm to resting-state functional connectivity to identify unique neural signatures of each WM component identified by CFA modelling, and examined whether the event cache had connectivity patterns distinct from those of the other WM components. Third, to further verify the functional implications of the event-associated brain regions identified with resting-state data, two established event tasks were conducted in the MRI scanner, and the activation and temporal dynamics of regions of interest (ROI) were examined.

To estimate the latent variable models of WM, we used 14 tasks to tap the different components of WM. There were four event storage tasks (point-light display biological movements, solid biological movements, movements of rectangles, and movements of circles), three object storage tasks (colors, shapes, and locations), two binding storage tasks (color-location and color-letter bindings), two CE tasks (Anti-saccade and N-back), and three other tasks (operation span [Ospan], symmetry span [Sspan], and multiple-object tracking [MOT]). The two complex span tasks (Ospan and Sspan) ^4,36,37 have been used frequently to tap the capacity of storing information in the presence of distractions, reflecting the joint function of storage and CE of WM. ³⁸ The MOT, which has dynamic features without higher-order stability, relies on the visuo-spatial sketchpad to process moving objects for a short interval. ^39–42

Descriptive statistics and correlations between the WM measures

Cowan’s K²⁰ was used to estimate the WM capacity for each storage task, the formula of Scholl et al. ⁴³ was used to estimate the tracked capacity of the MOT, and task accuracy was used to indicate the WM performance for the CE tasks and span tasks. Internal reliability for the tasks was calculated using either Cronbach’s alpha or the split-half correlation adjusted using the Spearman-Brown prophecy formula. ⁴⁴ Most measures showed acceptable levels of reliability (≥ .60, Supplementary Table S1) and relatively low skewness and kurtosis values (between − 2 and 2). ⁴⁵ The intercorrelations among all measures indicated that tasks assumed to tap the same WM component showed relatively higher correlations than those of a different WM component (Supplementary Table S2), suggesting both the convergent and discriminant validity of these measures.

Latent variable modelling based on Baddeley and Cowan’s WM models

We initially evaluated three models according to Baddeley and Cowan’s WM models (Fig. 2). It is worth noting that events in the current tests were probed using visual stimuli, and verbal encoding was suppressed; hence, we did not consider the phonological loop in the multi-component model of WM.

Model 1 was constructed to test whether all tasks could fit Cowan’s embedded-process model, ¹⁹ which included CE and FoA. According to this model, different types of stimuli are stored in FoA. Therefore, all tasks entailing the storage of visuo-spatial stimuli, binding, and events were assumed to load on a common FoA latent variable, while other tasks relating to CE loaded on the CE latent variable. Models 2 and 3 were constructed to test how well the tasks fit Baddeley’s multi-component model. ¹⁸ The only distinction between these two models was that the event storage tasks were loaded on a visuo-spatial sketchpad (Model 2) or episodic buffer (Model 3). Table 1 displays the fit statistics of all the models. Models 1–3 showed poor fitness according to the comparative fit index (CFI), suggesting that the two well-established models cannot fully account for the underlying structure of WM. In particular, the representations of events are probably not stored in FoA, conventional visuo-spatial sketchpad, or episodic buffer. It is possible that there is additional storage space for events.

Event cache as an independent component in WM

We further tested another set of models considering an independent event cache for holding events in WM (Fig. 2). We established Model 4 based on Cowan’s model, wherein the FoA was divided into event cache and object cache, with the former being responsible for storing events and the latter for static visual stimuli. Alternatively, we established Model 5 according to Baddeley’s model. In Model 5, an event cache was introduced in addition to a visuo-spatial sketchpad and episodic buffer. An inspection of the fit statistics in Table 1 indicates that both Models 4 and 5 showed an acceptable fit to the data. These results clearly demonstrate that there is an independent event cache underlying the WM structure.

The two types of events in this study have distinct properties. Biological movements of humans contain rich social information, whereas non-biological movements of physical shape do not. ^46,47 To further explore whether there are two different types of event cache owing to the difference in social semantics, we constructed Models 6 and 7 based on Models 4 and 5, respectively, by separating the event cache into biological movement and non-biological movement latent variables. Models 6 and 7 were acceptable. Model 6 showed a better fit to the data than the other three models according to the comparisons of CFI and Akaike’s information criterion (AIC) (Table 1).

Loading biological movement and non-biological movement into other factors worsens model fitting

We established several alternative models to test whether biological and non-biological movement factors can be integrated into other factors (Supplementary Figure S1). In Models 8–10, the non-biological movement tasks were loaded on the visuo-spatial sketchpad (Model 8), episodic buffer (Model 9), or object cache (Model 10). In Models 11–13, the biological movement tasks were loaded on a visuo-spatial sketchpad (Model 11), episodic buffer (Model 12), or object cache (Model 13). The results showed that all six models were unacceptable (Table 1). These results further suggest that events, including biological and non-biological movements, are represented separately from objects in the WM.

Hierarchical structure of WM subcomponents

Given that the tasks of biological and non-biological movements both tap the nature of events, we tested whether an underlying common event cache could be captured by a higher-order factor. We constructed Model 14 based on Model 6 (the best non-hierarchical model), where the event cache was conceptualized as a second-order latent variable extracted from biological- and nonbiological-movement factors (Fig. 3). This model had an adequate fit with the data that was even better than Model 6 according to AIC (Table 1), which indicated the hierarchical nature of the event cache in WM. Taken together, we consider Model 14 to be the best description of the WM construct.

Differentiation of event cache from the storage of motion

It is reasonable to conjecture that the event cache differed from the object cache in the model, mainly because the to-be-remembered stimuli were dynamic in the event storage tasks but static in the object storage tasks. To rule out this possibility, we constructed four models in which the MOT was loaded on the latent variable of biological movement (Model a), non-biological movement (Model b), object cache (Model c), and CE (Model d), based on Model 14. Table 1 presents the fit statistics of the models. The best model (Model c) shows that the object cache is responsible for storing moving objects. Therefore, the finding indicates that the event cache differing from the object cache is not due to the dynamic properties of task stimuli, but rather the intrinsically different mechanisms between event storage and object storage in WM.

In short, our behavioral data suggest that the event cache is an independent component of WM, which is distinct from the object cache that is responsible for the WM storage of single features and binding information. Moreover, this distinction is not due to the dynamic properties of the task stimuli.

Table 1

Fit statistics for working memory models
Models	AIC	χ²	df	RMSEA	CFI	SRMR
Model 1	7375.1	150.8	64	0.08	0.827	0.07
Model 2	7367.8	139.9	62	0.08	0.845	0.07
Model 3	7376.6	148.5	62	0.08	0.828	0.07
Model 4	7336.9	108.8	62	0.06	0.907	0.06
Model 5	7341.4	107.4	59	0.06	0.904	0.06
Model 6	7335.8	101.1	59	0.06	0.916	0.06
Model 7	7342.1	99.8	55	0.06	0.911	0.06
Model 8	7358.5	123.8	59	0.07	0.871	0.06
Model 9	7370.8	136.4	59	0.08	0.846	0.06
Model 10	7368.0	138.9	62	0.08	0.847	0.06
Model 11	7364.5	130.5	59	0.08	0.858	0.07
Model 12	7369.6	134.9	59	0.08	0.849	0.07
Model 13	7364.9	136.3	62	0.08	0.852	0.07
Model 14	7333.2	103.0	61	0.06	0.916	0.06
MOT
Model a	7868.0	126.3	73	0.06	0.902	0.06
Model b	7862.8	121.1	73	0.06	0.912	0.06
Model c	7855.3	114.0	73	0.05	0.925	0.06
Model d	7858.9	118.1	73	0.06	0.917	0.05
Note. RMSEA = root mean square error of approximation; CFI = comparative fit index; AIC = Akaike information criterion. Boldface indicates the best-fitting Model chosen.

Support vector regression prediction model of three WM components

Based on Model 14, we further applied SVR on resting-state functional connectivity to identify the neural signatures of each WM component (i.e., event cache [EVENT], object cache [OBJECT], and central executive [CE]). The functional connectivity matrix was computed for each participant using a brain atlas consisting of 268 areas (nodes) comprising the whole brain, ⁴⁸ which were assigned to 10 brain networks ⁴⁹ to facilitate network-based analysis (Fig. 1). Specifically, the time course of each node was extracted and a Pearson correlation coefficient was calculated between each pair of nodes, which was further transformed to Fisher’s Z score to ensure normality. An upper triangular connectivity matrix (35778 edges) was used in the prediction analyses. Based on previous studies that conducted SVR to reveal brain-behavior associations, a prediction model with a feature number of approximately 200 achieved good performance. ^33,50 Thus, we selected several subsets of features with approximately 200 features to investigate the prediction performance. Specifically, five subsets of features ranking from the top 1‰ to 9‰, with an increase step of 2‰, were used to predict the individual differences in WM components (with feature numbers ranging from 35 to 322, more subsets were also tested; see Supplementary Figure S2).

Prediction analyses revealed that all three WM components could be predicted by resting-state functional connectivity patterns (Fig. 4A, B, Supplementary Figure S2). In particular, the EVENT component could be predicted by a relatively small set of edges (with thresholds of 1‰ and 3‰). With a threshold of 1‰, the prediction model achieved the best performance, and the predicted EVENT scores were positively correlated with the actual values (r = 0.36, permutation p < 0.001; Fig. 4B, C). The model for OBJECT prediction achieved the best performance under the threshold of 9‰, and the predicted OBJECT scores were positively correlated with the actual values (r = 0.36, permutation p = 0.001, Fig. 4B, C). For the CE component, the prediction model could successfully predict individuals’ CE scores under thresholds of 5‰, 7‰, and 9‰, and the model performed the best under the threshold of 5‰. The correlation between predicted values and actual values was significant (r = 0.28, permutation p < 0.001, Fig. 4B, C). The characteristics of the selected features in the prediction models are visualized under the corresponding thresholds in the following sections.

The importance of different brain networks for the three WM components

To evaluate the importance of different brain networks in the prediction of these three WM components, the 268 nodes were grouped into 10 networks ⁴⁹ (Fig. 5, Supplementary Table S3), and the edges selected as features in the prediction models were plotted (Fig. 5A). As different features were selected in each iteration, only the features selected in all iterations (103) were considered, and the absolute weights of these features were summed to represent the connectivity between networks (Fig. 5A) with the calculation formula as follows: \({Edge}_{IJ}=\sum _{i\in I, j\in J}\left|{edge}_{ij}\right|\). Capital letters “I” and “J” refer to Network I and Network J, and “Edge” refers to the connectivity strength between two networks. Lowercase letters “‘i” and “j” refer to node i and node j, and “edge” refers to the feature weight between the two nodes. To quantitatively illustrate the relative contribution of each network, the relative degree (RD) of each network was calculated. The degree of each network (number of edges belonging to a network) was normalized by the total number of existing network degrees using the following formula: \({RD}_{I}=\frac{\sum _{J}{binarize(Edge}_{IJ})}{\sum _{I}\sum _{J}binarize({Edge}_{IJ)}}\). The capital letters refer to networks and the “binarize” function transfers a non-zero value to a 1 and keeps a zero-value unchanged. The RD values are shown in the radar graphs (Fig. 5B). According to the RD values (Fig. 5B, Supplementary Table S4), the cerebellum network had the highest RD among the ten networks in the EVENT prediction model and was considered the most important network for predicting the EVENT component, whereas the medial frontal, frontal-parietal-temporal associative, salience, and visual association networks were the top four networks with similar RD values for OBJECT prediction. The frontoparietal network and the default mode network played the most important roles in predicting the CE component.

The contributions of nodes to the prediction for three WM components

Although we identified the most important networks for each WM component, which brain regions within these networks contribute the most remained unclear. To this end, we calculated the degree of nodes (also referred to as node degree) within these networks and ranked nodes according to their degrees to identify the most important nodes within each network (Supplementary Figure S3, Supplementary Table S5; for node degree of all networks, please refer to Supplementary Figure S4). ⁵¹ The results displayed that the node with the largest degree was in the left posterior cerebellum, mainly in Crus I for the EVENT component. As for the OBJECT component, the nodes consisting of the left middle/superior frontal gyrus, right middle/superior temporal gyrus, left inferior/middle occipital gyrus, right post-central gyrus, superior parietal lobule and middle cingulate cortex contribute the most in prediction. In contrast, for the CE component, the right middle frontal gyrus within the frontoparietal network and the node covering the left middle occipital gyrus, middle temporal gyrus, angular gyrus, and precuneus within the default mode network had the largest degree.

Virtual lesion study on the role of cerebellum in prediction of EVENT

To further validate the essential role of the cerebellar network for the EVENT component, the prediction procedure was performed after removing nodes belonging to the cerebellar network. ⁵² Permutation tests were performed to confirm the significance, and a p value smaller than 0.01 was considered significant, controlling for multiple comparisons of five thresholds for each network. The results showed that, without the cerebellum, the EVENT component could not be successfully predicted (Fig. 6A), whereas the EVENT component could still be predicted when removing any other network (Fig. 6B). To further investigate whether the role of the cerebellar network is specific for EVENT prediction, we also performed virtual lesion analyses for CE and OBJECT, which showed that without the cerebellar network, CE and OBJECT could still be predicted (Fig. 6A), indicating that the cerebellar network was only essential in predicting the EVENT component. In addition, no single network played a decisive role in predicting OBJECT because the OBJECT component could be successfully predicted regardless of which network was removed (Supplementary Figure S5). For the CE component, the prediction model failed when the frontoparietal network, medial frontal network, or visual B network were removed (Supplementary Figure S6).

ROI-based analysis revealed the engagement of cerebellum in event WM tasks

Both the prediction analyses and virtual lesion analyses indicated that the resting-state connectivity of the cerebellum played an important role in predicting the EVENT component. Additionally, the assessment of nodal contribution to the prediction indicated that the node of the left posterior cerebellum Crus I was the most stable feature for prediction (Supplementary Figure S3). However, whether this region, identified by resting-state connectivity analysis, is functionally involved in WM processing is unclear. To address this issue, we further examined the activations (i.e., beta values) of this cerebellar node (Fig. 7A) using general linear modelling analysis (GLM) on an event N-back and a change detection fMRI task.

The results suggested that Crus I in the left posterior cerebellum was significantly activated in the N-back task (both at low and high loads for biological movement/non-biological movement conditions; Fig. 7B). Furthermore, significantly higher activation for the high load condition (2-back) relative to the low load condition (0-back) was found for both biological and non-biological movement conditions. Similarly, significant activation during the encoding phase of the change-detection task was found in this node (both under low and high loads for biological movement/non-biological movement conditions). Further analyses revealed significantly higher activation for the high load condition (4sets) relative to the low load condition (2sets) during both the encoding and delay phases (Fig. 7C, D). A time-series analysis also confirmed significantly higher activations for the high load condition (4sets) relative to the low load condition (2sets) during the encoding and delay phases in both biological and non-biological movement tasks, although the activation was lower in the delay phase than in the encoding phase (Fig. 7E). In short, using two different stimulus materials with two different WM paradigms, our data showed that the cerebellar node with the highest predictive power was consistently activated in a load-dependent manner during WM tasks requiring the involvement of the EVENT component.

By using biological and non-biological movements as the representative stimuli of events in daily life, our study, for the first time, revealed an independent storage component, that is, the event cache, in WM. The event cache, as a psychological construct, is distinct from the WM storage space holding objects and related features (i.e., the object cache in the current study). Moreover, SVR prediction analyses based on resting-state functional connectivity successfully predicted event cache, object cache, and CE components with distinct patterns. Critically, the cerebellar network was found to be the most essential network for the event cache, and the left cerebellum Crus I displayed the largest node degree in the prediction model. This result provided convergent evidence from neural data to support the event cache as an independent component of WM. While the main body of WM studies has focused on objects and their constituent features, our study strongly suggests the necessity and importance of extending the scope of WM exploration from objects and features to events.

Event cache as an independent psychological construct of WM

The demonstration of an event cache in WM highlights two key theoretical issues. First, we revealed a new independent component in WM. The independence of the event cache is not merely driven by the dynamic features of the stimuli, as the storage of moving objects was found in the object cache, which aligns with previous studies. ^39,53,54 The identification of an event cache supports a set of previous empirical findings showing that human actions are stored independently from colors, shapes, locations, and binding representations. ^55–58 It also supports Baddeley’s speculation that there might be an independent store holding human movements in WM. ² Critically, the event cache may be a hierarchical construct consisting of two subcomponents: biological and non-biological movements. This further dissociation is congruent with recent findings that only WM of biological movements has an intimate relationship with empathy and theory of mind. ^46,47 It is worth noting that distinct mechanisms exist between different hierarchies, according to the emergence principle of Anderson. ³² Therefore, although we have known much about WM mechanisms of object in the past 40 years, we need to pay attention to the mechanisms of events in WM in the future. For example, it remains unclear how events are rehearsed in WM and how events in WM guide external attention. However, since there are many different types of events and we only tested events in vision, future studies should consider how events of other modalities (e.g., verbal events) are retained in WM.

Second, our findings imply that event cache is responsible for the storage space for the product of event segmentation. Previous event studies have implied that after event segmentation, the resultant events were retained in a multimodal episodic buffer. ²⁴ However, we did not find evidence supporting an independent episodic buffer in WM. Although all stimuli were visually presented in the tasks, we found that visual events had an independent storage space in WM in addition to the object cache. Therefore, the visual events, theoretically, should be stored in the event cache in WM. Taking this into consideration, we argue that two important questions should be addressed in the future: how the object cache and event cache interact to form a coherent experience and how events containing multimodal information are built in WM.

Both visual features and bindings are retained in object cache

Our best model supports the existence of an object cache that is mainly responsible for the storage of visual objects and their constituent features. In contrast to our object cache, Baddeley’s model suggests that the information of bindings and their constituent features are stored in episodic buffer and visuo-spatial sketchpad, respectively. The episodic buffer was initially assumed to actively bind information from different sources into one episode with the help of CE. ¹⁸ Later empirical studies implied that the binding process in the episodic buffer was passive. ⁵⁹ However, few studies have directly examined whether an independent component exists in handling binding. To the best of our knowledge, only two CFA studies have explored whether the episodic buffer is an independent component in WM, but with contradictory conclusions. ^60,61 One potential caveat is that both studies relied on children whose WM remains developing. Our study was the first to explore the independence of binding-related psychological constructs in healthy adults.

Our findings suggest that although the to-be-remembered representations in the binding and feature tasks are different, they are processed by the same object cache, and hence, may share similar mechanisms. Indeed, recent empirical findings suggest that object-based attention plays a key role in retaining both bindings and single-featured objects in WM. ^62–64 Therefore, we argue that episodic buffer as an independent buffer for handling bindings should be reconsidered. Note that, although color-letter binding theoretically tapped the binding between visual and verbal codes, participants may process the letter both visually and verbally in a short maintenance period; ⁶⁵ hence, color-letter bindings in our setting have visual-related representations. Since all bindings in our study contained visual information, future studies need to test the generalization of the current findings using bindings of other modalities (e.g., auditory). For example, there may be an auditory-related binding component underlying the phonological loop and an event-related binding component underlying the event cache.

Event cache, object cache, and CE have distinct neural correlates

We used resting-state functional connectivity patterns to predict WM component scores extracted from the corresponding WM tasks. These results suggest that the event cache has a unique neural substrate and should be separated from the object cache. Although the cerebellum has been found to be activated in previous WM studies, ^66,67 it is a novel finding that the cerebellum plays the most essential role in predicting the event cache component. This may be attributed to the application of the connectivity-based method, which weighs more on the inter-region connection rather than the local activation of a region.

Traditionally, the cerebellum is considered important for motor coordination and learning. ^68–71 However, cumulative evidence supports the involvement of the cerebellum in various high-level cognitive functions, including WM. ^72,73 In particular, the posterior cerebellum, which includes lobule Crus I which was predictive of event cache in our study, was significantly activated in an N-back task. ⁶⁷ Our ROI analysis also showed activation of the Crus I node in response to both biological and non-biological movements. Relatedly, the posterior cerebellum (especially in Crus I and II) was found to be closely associated with the ability to reconstruct action sequences in a chronological order. ^74–78 According to a hierarchical processing perspective of WM, ^22,66,79 it could be speculated that the cerebellum is involved in constructing high-order representations (referred to as events in our study) of what happened to person(s) or object(s) from dynamic stimuli. Corroborating this speculation, the MOT task does not load on the event cache in our model. Though the MOT task is composed of dynamic visual stimuli, participants only need to track/update the targets’ locations rather than integrate all the moving steps to understand what happened to the targets. Our results not only provide the first evidence to support that the event cache should be considered as a WM component that processes sequential stimuli with temporal structures, but also provide empirical neural evidence that the cerebellum Crus I may serve as one of the primary neural underpinnings.

As for the object cache component, previous studies have identified that the frontal, parietal, and occipital regions are significantly activated during color, shape, and color-shape binding WM tasks ^22,80,81 and after applying anti-phase gamma transcranial alternating current stimulation (tACS) to the temporal and parietal cortices, the color-shape binding WM performance was significantly increased in low-performers. ⁸² Our finding that the medial frontal, frontal-parietal-temporal associative, salience, and visual association networks played vital roles in the prediction model is consistent with these results. In addition, multivariate pattern analyses have identified that brain regions related to the maintenance of object features spread across almost the entire cerebral cortex, ^22,81 which also converged with our virtual lesion results showing that no single brain network plays an absolutely dominant role in object maintenance.

As for the CE component, we found that the frontoparietal network and the default mode network contributed the most to the CE component prediction, and the connectivity between the two networks also contributed to the prediction. The frontoparietal network has consistently been reported as the core neural substrate for tasks requiring executive control for information integration or manipulation. ^66,83−86 Moreover, the strength of the functional interaction between the frontoparietal network and the default mode network was associated with individual’s performance on the N-back task, ⁸⁷ suggesting that the resting-state connectivity-based prediction method can recapture these neural pathways reported by studies that only examined a few sets of connections. Notably, in virtual lesion analyses, we found that the CE component could not be successfully predicted without the frontoparietal network. This result not only supports previous findings by showing that the frontoparietal network is an indispensable neural underpinning supporting CE, but also demonstrates the feasibility of identifying important neural correlates of the event cache and object cache components using the same procedure.

Participants and experimental designs

A total of 208 participants (150 females, 58 males) were recruited from the Zhejiang University and Hangzhou Normal University. Two participants who failed to complete all experiments were excluded. The final sample that completed behavioral tests (including seven questionnaires and fourteen computer versions of WM tasks) consisted of 206 participants (148 females, 58 males) between the ages of 17 and 25 (M = 20.4, SD = 1.4). All participants had normal color vision and normal or corrected-to-normal visual acuity. They provided written informed consent prior to the experiments and received payment or course credit for their participation. One hundred and seven participants among 206 participants further completed MRI experiments, which included high-resolution T1-weighted scanning, resting-state fMRI scanning, and task fMRI scanning (part of the task fMRI data were reported in another study ⁶⁶). As for resting-state fMRI analysis, four participants were excluded because of large head motion (maximum displacement > 3 mm or 3◦, or the number of time points with Framewise Displacement > 0.5 mm exceeded 15% of the total), resulting in 103 participants in the final fMRI sample (M = 19.5, SD = 1.3; 61 females, 42 males). For the ROI analysis, a subgroup of 47 participants (M = 19.9, SD = 0.8; 29 females, 18 males) among 107 participants was included to perform the N-back task and another 34 participants (M = 23.3, SD = 2.6; 16 females, 18 males) were recruited to perform the change detection task independently (for details of N-back and change detection task, please refer to Zhou et al. ⁶⁶). This study was approved by the ethics committee of Zhejiang University.

The total duration of the behavioral experiments was approximately 210 minutes. To control for the possible influence of the fatigue effect, we divided the testing into three sessions, which were conducted separately for three days. In Session 1, participants completed seven questionnaires that were irrelevant to our study and were not reported here. In Sessions 2 and 3, they completed a series of WM tasks in a dark compartment of the laboratory. Resting-state fMRI data and the N-back task fMRI data were collected after the participants completed all behavioral tests. An independent group performed an event change detection task in the fMRI scanner (preliminary results have been reported in a previous study ⁶⁶).

WM tasks

The stimuli in the 14 WM tasks were generated using MATLAB Psychophysics Toolbox and presented on a black (0, 0, 0) background. Participants sat in a dark room, 60 cm in front of a 17-inch CRT monitor with a resolution of 1024 × 768.

For storage tasks (event storage, object storage, and binding storage), the change-detection paradigm was adopted, and the number of memory items was set to two, four, and six (except for the location storage task, which was set to four, six, and eight). There were 26 trials under each memory load, and 78 trials in each task. Participants were allowed to take a break after every 26 trials. Before the formal experiment, an 8-trial practice was conducted.

Event storage tasks

Event storage tasks included biological movement (BM) and non-biological movement (NBM) tasks. Point-light displays (PLDs) and solid agents were used as two types of BM stimuli, and rectangular and circular movements were used as two types of NBM stimuli.

PLD BM. For each PLDs movement, 13 light points were placed at the distinct joints of a moving human body to form a coherent and meaningful movement. We selected nine movements from the database of Vanrie and Verfaillie ⁸⁸: cycling, jumping, painting, spading, walking, waving, chopping, paddling, and saluting (see Supplementary Figure S7 B). Every animation consisted of 30 distinct frames, and each frame was displayed twice in succession, leading to a 1-s PLD (refresh rate, 60 Hz). Each stimulus subtended a visual angle of approximately 1.64° × 1.64°. The spatial locations of the PLDs were randomly selected from eight evenly distributed spots on an invisible circle with a radius of 4.88° from the screen center.

The procedure of a single trial was shown in Figure S7 A. Each trial began with a display of the word "Coca-Cola" (in Chinese) for 500 milliseconds (ms), which reminded participants to rehearse aloud during the whole task. This manipulation was set to prevent the participants from verbally coding the BM. After an interval of 150–350 ms, the memory array appeared for four seconds. Participants were required to remember these stimuli. After a 900-ms retention interval, a red probe was presented at the center of the screen. Participants were required to determine within 3 seconds (s) whether the red probe had appeared in the memory array. They pressed the "J" key if it had appeared in the memory array, and pressed the "F" key if it had not. The probe BM was the same in 50% of the trials and changed to a new action in the remaining 50% of the trials. After the response, there was an interval of 500–700 ms between trials.

Solid BM. Solid agents were used as the other type of BM stimuli. In this task, BMs were presented in the form of solid agents, which were created using the Poser software. We selected nine movements: arm, foot and hip, jumping, squatting, walking, waving, saluting, turning, and bowing (see Figure S7 C). The other settings were the same as those in the PLD BM task.

Rectangle movement (Rec Move). The movements of the rectangles were created using a 12-dotted PLD. ⁴⁷ In line with the PLD BM stimuli, the dotted rectangle showed nine distinct movements (see Figure S7 D). (1) The left and right sides moved downward by 60° relative to their vertical positions and then returned. (2) The left side moved upward by 45° relative to its vertical position and then returned, while the right half moved upward by 45° relative to its vertical position and then returned. (3) The left and right halves moved downward by 90° relative to their vertical positions and then returned, while the top side rotated around the middle dot once. (4) The top side moved upward by 90° relative to its horizontal position and then returned. (5) The top half of the rectangle rotated 90° clockwise relative to its vertical position, while the left bottom half moved upward by 90°. (6) The top side rotated around the top dot on the right side by 180°, while both the left and bottom sides moved downward by 90°. (7) The right side moved downward by 45° and then returned. (8) The left side rotated 180° clockwise around the middle dot, while the right side moved downward by 90° relative to its vertical position. (9) The top side moved downward by 90° relative to its horizontal position and then returned, while the right side moved 180° clockwise relative to its horizontal position and then returned. The other settings were the same as those in the PLD BM task.

Circle movement (Circle Move). We used 13-dotted PLDs of circular movement, containing nine distinct movements ⁸⁹: moving up, moving down, rolling right, rolling left, shrinking, inflating, moving diagonally, splitting horizontally, and splitting vertically (see Figure S7 E). The initial radius of each circle is approximately 1.15°. Each movement lasted for 1 s. The other settings were the same as those in the PLD BM task.

Object storage tasks

Color. For distinct colors (Figure S7 F), we used white (255,255,255, in RGB value), yellow (255,255,0), lime (0,255,0), gray (128,128,128), light pink (255,178,193), aqua (0,255,255), blue (0,0,255), red (255,0,0), and magenta (255,0,255). The spatial locations of the colored squares (1.1° × 1.1°) were distributed at eight spots on an invisible circle with a radius of 4° from the center of the screen. After a 300–400 ms interval, the memory array was displayed for 200 ms. The other settings were the same as those in the PLD BM task.

Shape. We replaced the nine colored squares with nine distinct shapes (see Figure S7 G; 1.6° × 1.6°). The other aspects were the same as those of the color WM task.

Location. Participants were required to remember the locations of four, six, or eight white squares (0.2° × 0.2°) to avoid the ceiling effect. They judged whether the probe (red square) appeared at one of the memorized locations. The stimuli appeared inside an invisible circle with a radius of 5° from the center of the screen. These squares were evenly distributed in the four quadrants and did not overlap. A fixation (“+”) remained at the center of the screen from the blank interval after the Coca-Cola presentation until the end of the detection phase. The other settings were the same as those in the color task.

Binding tasks

Binding tasks were designed to assess WM capacity when two different features were presented within the visuo-spatial domains (color-location binding) or across the verbal and visual WM domains (color-letter binding). Participants had to hold them together in WM to respond correctly to the task. The probe could be one of the old items in the memory array in 50% of trials. In the remaining trials, the probe was formed by combining the different dimensions of the two items in the memory array.

Color-location binding. Participants were required to remember the bindings between the color and location of the 0.8° × 0.8° squares. The same colors were chosen from the color task, except that light pink was replaced with green (0,128,0). The squares appeared for 200 ms in an area of 10° × 10° in the screen center and were non-overlapping. Subsequently, a probe appeared at a certain location on the screen, and participants were required to report whether its color matched the memory item at the corresponding location. The fixation settings were the same as those used for the location task.

Color-letter binding. Participants were required to remember the binding between color and capital letters (A, B, C, D, E, H, M, J, and K). Each letter subtended a visual angle of approximately 1.3°×1.3° and was presented on an invisible circle with a radius of 4° from the center of the screen for 200 ms. A colored capital letter then appeared at the center of the screen. Participants were asked to determine whether any memory item matching both letter and color appeared. No fixations were demonstrated during the task.

Central executive tasks

Anti-saccade (Anti). This task was adapted from the task used by Unsworth and Spillers. ⁹⁰ Each trial started with a fixation point displayed for 200, 600, 1000, 1400, or 1800 ms. A cue was then displayed either to the left or right of the fixation point for 100 ms, followed by a 50-ms interval. This procedure was repeated a second time. After the disappearance of the cue, a target letter (B, P, or R) was immediately presented to the opposite side of the second cue for 100 ms, which was followed by a masking stimulus (an H for 50 ms and then an ‘‘8’’ that remained on screen until a response). Participants were required to identify the target letter (B, P, or R) by pressing a corresponding key (1, 2, or 3). There were 15 practice and 40 formal trials. The proportion of correct responses was then recorded.

N-back. The task began by showing the fixation in the center of the screen for 1000 ms. A sequence of black squares then appeared in the eight outer squares of a three-by-three grid (the middle square was not used), and participants were asked to decide whether the location of each square matched the one appearing three items (3-back) before. The presentation time was 500 ms for each square, with inter-stimulus fixations of 2000 ms. Participants pressed “F” if the locations were identical or “J” if they were not. The task included two blocks, each with 40 formal trials. We conducted eight practice trials before the formal experiment. The proportion of correct responses was then recorded.

Other WM tasks

Operation span (Ospan). Participants remembered a set of consonants (F, H, J, K, L, N, P, Q, R, S, T, and Y) while solving arithmetic problems. Three practice sessions preceded the formal trials to familiarize participants with the procedure. The first session was a simple letter span task that asked participants to recall a set of consonants in the same order as they were presented. Each letter was presented in the center of the screen for 1000 ms. During recall, a 4 × 3 matrix of letters was displayed, and participants had to select the letters in the correct order by clicking the box beside the letter. In the second session, the participants were asked to solve 15 arithmetic problems (e.g., (3 × 2) + 6 =?). The processing time for each problem was recorded. After completing all problems, the program calculated the mean and standard deviation (SD) of the time each participant took. The third session combined the simple span and arithmetic tasks. Participants were first presented with a mathematical operation and then a letter. The time limit for each operation was fixed to the average time plus 2.5 SD obtained in the second session so that participants had little time to rehearse the letters. After the participants completed all three practice sessions, the program proceeded to the formal trials, which were similar to the trials in the third practice session. The list length varied randomly from four to eight letters. Three trials were performed for each set size. The dependent variable was the number of correct lists that recalled all letters in the correct order. Participants repeated the task if the overall accuracy of the math portion was lower than 80%.

Symmetry span (Sspan). Participants were required to remember the location of sequences of red squares (1.1° × 1.1°) presented within a 3 × 3 matrix while performing a symmetry judgment task. In line with Ospan, the task started with three practice sessions: two trials of the simple square-span task, 15 symmetry judgment tasks, and three trials of the combined task. In the formal trials, participants first saw a black-and-white 8 × 8 grid pattern and decided whether it was symmetrical along the vertical axis. The time limit was calculated in the same manner as for Ospan. Then, participants were presented with a 4 × 4 matrix with one of the cells filled in red for 650 ms. During recall, a blank 4 × 4 matrix appeared, and participants had to recall the red squares in the same order as they were presented by clicking on the corresponding locations. There were three trials for each set size with a list length ranging from three to six. The same scoring procedure was used as with the Ospan task.

Multiple-object tracking (MOT). Twelve white disks with a viewing angle of 0.25° were presented in an area of 8.5° × 8.5° at the center of the screen. The movement of the disks was divided into three stages: marking, tracking, and detection. The marking phase lasted for 3000 ms, during which all disks remained still. However, two, four, or six disks turned red and solid for 2500 ms, prompting the participants to track the position of these disks. Subsequently, all disks became hollow and appeared for 500 ms. The tracking phase lasted for 4500 ms. All discs moved in random directions at speeds ranging from − 0.00058°/ms to 0.0064°/ms, and they did not overlap during the movement. The detection phase lasted for a maximum of 1500 ms. All discs stopped moving and one of the discs turned solid green. Participants were required to determine whether the disc was one of several discs marked in the marking phase. This task did not require verbal suppression, and the other aspects were the same as those in the PLD BM task.

Behavioral analysis

For WM storage tasks, the WM capacity for each type of stimulus was estimated using Cowan’s formula ⁹¹: K = S × (H − F), where K is the WM capacity, S is the number of to-be-memorized stimuli, H is the hit rate, and F is the false alarm rate. We calculated K for each set size for each participant and considered K_max among the three load conditions as one’s WM capacity.^56,92,93 For all tasks, univariate outliers were defined as individual scores that exceeded 3 SDs from the respective grand mean. Of a total of 2884 observations, five met this criterion. These scores were replaced with corresponding cut-off values (M ± 3SD). For the MOT task, we used the formula of Scholl et al. ⁴³ to derive the effective number of objects tracked: K = S × (2P − 1), where K is the effective number of objects tracked, S is the number of targets, and P is the tracked accuracy in a tracked-load condition.

All structural equation models were estimated using Mplus 8. ⁹⁴ The robust maximum likelihood estimation method (MLM) that has been developed for non-normal data was used in modelling estimation. ⁹⁵ The evaluation of the fit statistics was based on the criteria recommended by Kline ⁴⁵ and Distefano ⁹⁶. Specifically, the fit of a model was considered good (or acceptable) if normed χ2 (χ2/df) ≤ 2 (3), RMSEA ≤ 0.05 (0.08), SRMR ≤ 0.05 (0.10), and CFI ≥ 0.95 (0.90). We made model comparisons by considering changes in CFI and AIC. A difference of .01 or larger in CFI was considered a substantial difference. ⁹⁷ AIC was used to compare non-nested models, in which a model with a smaller AIC was preferred. ⁹⁸

MRI imaging parameters and pre-processing

MRI data were acquired on a 3T Siemens Prisma scanner using a 20-channel coil at the Center for Brain Imaging Science and Technology, Zhejiang University. A T2*-weighted single-shot echo-planar imaging sequence with multiband acceleration (multiband factor = 4) was used to collect resting-state and task functional images. The participants were instructed to look at a fixation and remain stationary during the resting-state scanning. The acquisition parameters were as follows: TR/TE = 1000/34 ms, flip angle = 50°, FOV = 230 × 230 mm², matrix = 92×92, voxel size = 2.5×2.5×2.5 mm³, and slice number = 52. High-resolution T1-weighted anatomical images were collected using a T1-weighted 3D magnetization-prepared rapid gradient echo sequence with the following parameters: TR/TE = 2300/2.32 ms; flip angle = 8°; FOV = 240 × 240 mm²; matrix = 256 × 256; voxel size = 0.94 × 0.94 × 0.9 mm³; and 208 slices in the sagittal panel. Resting-state fMRI data were pre-processed using AFNI, ⁹⁹ ANTs (http://stnava.github.io/ANTs/), SPM12 (https://www.fil.ion.ucl.ac.uk/spm/), and DPABI. ¹⁰⁰ The first five volumes were discarded to allow for signal equilibrium, and the remaining volumes were first corrected for slice acquisition timing and then for head motion using rigid-body transformation. Spatial normalization was further applied to the fMRI data, and spatial smoothing was conducted with a five mm full-width-at-half-maximum Gaussian kernel. To further control for head motion influence, the six motion parameters and their temporal derivatives, as well as the first five principal components extracted from the white matter and CSF regions, were regressed out from the smoothed data, and finally, band pass filtering (0.01 Hz-0.1 Hz) was applied.

Brain-behavior prediction analyses

We calculated the scores of all latent variables to investigate the neural correlates underlying different WM components. We first transformed the scores (accuracies for CE tasks and K_maxs for storage tasks) of each WM task into z-scores. According to the structure of Model 14, we then used the average of the z-scores of the corresponding tasks to attain the score of each first-order factor. For example, BM = 0.5 * (zPLDBM + zSolidBM). The score of the second-order factor is averaged over the scores of the first-order factors; for example, EVENT = 0.5 * (BM + NBM). Thus, we obtained scores for the EVENT, OBJECT, and CE components. SVR models were trained using the leave-one-out cross-validation (LOOCV) method to predict the scores of the CE, EVENT, and OBJECT components derived from factor analysis. The prediction procedure was performed using the MVPANI toolbox, ¹⁰¹ which was programmed based on the LIBSVM toolbox (http://www.csie.ntu.edu.tw/~cjlin/libsvm/). Feature ranking and selection were performed using a training dataset. For each iteration, one participant was left out as the testing sample, whereas the remaining participants constituted the training dataset. Within the training dataset, the predictive weight of each edge was determined through SVR model training, and the absolute values of the weights were sorted in descending order. Based on previous studies that conducted SVR to reveal brain-behavior associations, a prediction model with a feature number of approximately 200 achieved good performance. ^33,50 Thus, five subsets of features ranking from the top 1‰ to 9‰, with an increase step of 2‰, were used to predict the WM score of the testing sample in the current study (with feature numbers ranging from 35 to 322, more subsets were also tested; see Supplementary materials). The correlation coefficient between the predicted WM score and the actual score was calculated to quantify the prediction accuracy with permutation to test the significance of this correlation. For each permutation, the WM scores were shuffled and the same LOOCV prediction procedure was performed to generate a correlation coefficient between the predicted and shuffled WM scores. After 1000 permutations, the significance p value was calculated as the percentage of correlation coefficients exceeding the one calculated without shuffling WM scores. A p value smaller than 0.01 was considered significant, controlling for multiple comparisons of five threshold values for each WM component.

ROI analyses of event WM tasks

An event N-back task and a change detection task were employed to investigate whether the brain regions identified as important by the prediction models participated in WM processing. Both tasks used PLD biological and non-biological movements as stimuli and consisted of two WM loads (0back and 2back for the N-back task, 2sets and 4sets for the change detection task). For the N-back task, biological movement 0back, biological movement 2back, non-biological movement 0back, and non-biological movement 2back were regarded as separate regressors in the general linear modelling (GLM) analyses. For the change detection task, biological movement 2sets, biological movement 4sets, non-biological movement 2sets, and non-biological movement 4sets regressors were constructed for both encoding and delay periods (for task and GLM details, please refer to Zhou et al. ⁶⁶). Beta values were extracted from each load for each stimulus type of the event WM task. We conducted one-sample t-tests to test the significance of activation, and paired t-tests to compare activation differences between low and high loads for each stimulus type (biological and non-biological movements).

Additionally, we conducted time-series analyses for the change-detection task to investigate the BOLD signal change percentage during the task. Specifically, for each participant, we extracted the average time-course of the ROI and segmented it according to the onset time of each encoding period. We then averaged the time-course segments of each load for each stimulus type. The resultant time-courses were further converted to the percent signal change for each condition by subtracting the value of the corresponding time points of the baseline trial and then dividing by the baseline trial value. We observed two peaks in the time-series which corresponded to the encoding and probe periods. Accordingly, we selected signals between 7 and 9 s for the encoding phase, signals between 17 and 19 s for the probe phase, and signals between 13 and 15 s for the delay phase. ¹⁰² Finally, we conducted paired t-tests to test the load effects (4sets > 2sets) for each stimulus type.

Acknowledgment

This research was supported by National Natural Science Foundation of China (32271090; 32000761; 81971245), Key Program of Natural Science Foundation of Zhejiang Province (LZ20C090001), Fundamental Research Funds for the Central Universities (No.2022QZJH09), Research of Basic Discipline for the 2.0 Base of Top-notch Students Training Program, the Ministry of Education of China (20211033), the Zhejiang Province ‘Qianjiang Talent Program’, the MOE Frontiers Science Center for Brain Science & Brain-Machine Integration, Zhejiang University, and the STI 2030-Major Projects (No. 2021ZD0200409).

Baddeley, A. D. & Hitch, G. Working memory, in Psychol. Learn. Motiv. Vol. 8 (ed Gordon H. Bower) 47-89 (Academic Press, 1974).
Baddeley, A. Working memory: theories, models, and controversies. Annu. Rev. Psychol. 63, 1-29, (2012).
Cowan, N. An embedded-processes model of working memory, in Models of working memory: Mechanisms of active maintenance and executive control (eds Atsushi Miyake & Patricia Shah) 62-101 (Cambridge University Press, 1999).
Unsworth, N., Fukuda, K., Awh, E. & Vogel, E. K. Working memory and fluid intelligence: Capacity, attention control, and secondary memory retrieval. Cognitive psychology 71, 1-26, (2014).
Jaeggi, S. M., Buschkuehl, M., Jonides, J. & Perrig, W. J. Improving fluid intelligence with training on working memory. Proc. Natl. Acad. Sci. 105, 6829-6833, (2008).
McVay, J. C. & Kane, M. J. Why does working memory capacity predict variation in reading comprehension? On the influence of mind wandering and executive attention. Journal of experimental psychology: general 141, 302, (2012).
Kellogg, R. T., Turner, C. E., Whiteford, A. P. & Mertens, A. The role of working memory in planning and generating written sentences. Journal of Writing Research 7, 397-416, (2016).
Lee, C. S. & Therriault, D. J. The cognitive underpinnings of creative thought: A latent variable analysis exploring the roles of intelligence and working memory in three creative thinking processes. Intelligence 41, 306-320, (2013).
Kumar, S. et al. Extent of dorsolateral prefrontal cortex plasticity and its association with working memory in patients with Alzheimer disease. JAMA psychiatry 74, 1266-1274, (2017).
Wang, Y. et al. A meta-analysis of working memory impairments in autism spectrum disorders. Neuropsychol.Rev. 27, 46-61, (2017).
Murphy, O. et al. Individuals with depression display abnormal modulation of neural oscillatory activity during working memory encoding and maintenance. Biol. Psychol. 148, 107766, (2019).
Reinhart, R. M. G. & Nguyen, J. A. Working memory revived in older adults by synchronizing rhythmic brain circuits. Nat Neurosci 22, 820-827, (2019).
Cowan, N. Working memory development: A 50-year assessment of research and underlying theories. Cognition 224, 105075, (2022).
Ludyga, S., Gerber, M. & Kamijo, K. Exercise types and working memory components during development. Trends Cogn. Sci. 22, 191-203, (2022).
Oberauer, K. Is rehearsal an effective maintenance strategy for working memory? Trends Cogn. Sci. 23, 798-809, (2019).
Baddeley, A., Hitch, G. & Allen, R. A multicomponent model of working memory, in Working memory: State of the science (eds Robert Logie, Valerie Camos, & Nelson Cowan) 10-43 (Oxford University Press, 2021).
Xie, W. et al. The medial temporal lobe supports the quality of visual short-term memory representation. Nature Human Behaviour, 1-15, (2023).
Baddeley, A. The episodic buffer: a new component of working memory? Trends Cogn. Sci. 4, 417-423, (2000).
Cowan, N. Evolving conceptions of memory storage, selective attention, and their mutual constraints within the human information-processing system. Psychological bulletin 104, 163, (1988).
Cowan, N. Processing limits of selective attention and working memory: Potential implications for interpreting. Interpreting 5, 117-146, (2000).
Oberauer, K. Access to information in working memory: exploring the focus of attention. Journal of Experimental Psychology: Learning, Memory, and Cognition 28, 411, (2002).
Christophel, T. B., Klink, P. C., Spitzer, B., Roelfsema, P. R. & Haynes, J. D. The Distributed Nature of Working Memory. Trends Cogn. Sci. 21, 111-124, (2017).
Shipley, T. F. & Zacks, J. M. Understanding events: From perception to action. (Oxford University Press, 2008).
Radvansky, G. A. & Zacks, J. M. Event cognition. (Oxford University Press, 2014).
Pavlova, M. A. Biological motion processing as a hallmark of social cognition. Cereb. Cortex 22, 981-995, (2012).
Johansson, G. Visual perception of biological motion and a model for its analysis. Perception & psychophysics 14, 201-211, (1973).
Shin, Y. S. & DuBrow, S. Structuring memory through inference‐based event segmentation. Topics in Cognitive Science 13, 106-127, (2021).
Zacks, J. M. Event perception and memory. Annu. Rev. Psychol. 71, 165-191, (2020).
Radvansky, G. A. Event segmentation as a working memory process. (2017).
Richmond, L. L., Gold, D. A. & Zacks, J. M. Event perception: Translations and applications. Journal of Applied Research in Memory and Cognition 6, 111-120, (2017).
Radvansky, G. A. & Zacks, J. M. Event boundaries in memory and cognition. Current opinion in behavioral sciences 17, 133-140, (2017).
Anderson, P. W. More is different: broken symmetry and the nature of the hierarchical structure of science. Science 177, 393-396, (1972).
Dosenbach, N. U. et al. Prediction of individual brain maturity using fMRI. Science 329, 1358-1361, (2010).
Yoo, K. et al. A brain-based general measure of attention. Nat Hum Behav 6, 782-795, (2022).
Yamashita, M. et al. A prediction model of working memory across health and psychiatric disease using whole-brain functional connectivity. Elife 7, (2018).
Unsworth, N., Heitz, R. P., Schrock, J. C. & Engle, R. W. An automated version of the operation span task. Behavior research methods 37, 498-505, (2005).
Li, C., Ren, X., Schweizer, K. & Wang, T. Strategy use moderates the relation between working memory capacity and fluid intelligence: A combined approach. Intelligence 91, 101627, (2022).
Engle, R. W. Working memory capacity as executive attention. Current directions in psychological science 11, 19-23, (2002).
Drew, T. & Vogel, E. K. Neural measures of individual differences in selecting and tracking multiple moving objects. J. Neurosci. 28, 4183-4191, (2008).
Balaban, H., Drew, T. & Luria, R. Delineating resetting and updating in visual working memory based on the object-to-representation correspondence. Neuropsychologia 113, 85-94, (2018).
Drew, T., Horowitz, T. S., Wolfe, J. M. & Vogel, E. K. Neural measures of dynamic changes in attentive tracking load. J. Cognit. Neurosci. 24, 440-450, (2012).
Luria, R. & Vogel, E. K. Come together, right now: Dynamic overwriting of an object's history through common fate. J. Cognit. Neurosci. 26, 1819-1828, (2014).
Scholl, B. J., Pylyshyn, Z. W. & Feldman, J. What is a visual object? Evidence from target merging in multiple object tracking. Cognition 80, 159-177, (2001).
Shipstead, Z., Lindsey, D. R. B., Marshall, R. L. & Engle, R. W. The mechanisms of working memory capacity: Primary memory, secondary memory, and attention control. Journal of Memory and Language 72, 116-141, (2014).
Kline, R. B. Principles and practice of structural equation modeling. (Guilford publications, 2015).
He, J., Guo, D., Zhai, S., Shen, M. & Gao, Z. Development of social working memory in preschoolers and its relation to theory of mind. Child Dev. 90, 1319-1332, (2019).
Gao, Z., Ye, T., Shen, M. & Perry, A. Working memory capacity of biological movements predicts empathy traits. Psychonomic Bulletin & Review 23, 468-475, (2016).
Shen, X., Tokoglu, F., Papademetris, X. & Constable, R. T. Groupwise whole-brain parcellation from resting-state fMRI data for network node identification. Neuroimage 82, 403-415, (2013).
Greene, A. S., Gao, S., Scheinost, D. & Constable, R. T. Task-induced brain state manipulation improves prediction of individual traits. Nat Commun 9, 2807, (2018).
Smyser, C. D. et al. Prediction of brain maturity in infants using machine-learning algorithms. NeuroImage 136, 1-9, (2016).
Feng, J. et al. A cognitive neurogenetic approach to uncovering the structure of executive functions. Nat Commun 13, 4588, (2022).
Feng, C. et al. Individualized prediction of trait narcissism from whole-brain resting-state functional connectivity. Hum. Brain Mapp. 39, 3701-3712, (2018).
Culham, J. C., Cavanagh, P. & Kanwisher, N. G. Attention response functions: characterizing brain areas using fMRI activation during parametric variations of attentional load. Neuron 32, 737-745, (2001).
Jovicich, J. et al. Brain areas specific for attentional load in a motion-tracking task. J. Cognit. Neurosci. 13, 1048-1058, (2001).
Liu, Y., Lu, X., Wu, F., Shen, M. & Gao, Z. Biological motion is stored independently from bound representation in working memory. Visual Cognition 27, 701-713, (2019).
Shen, M., Gao, Z., Ding, X., Zhou, B. & Huang, X. Holding biological motion information in working memory. J Exp Psychol Hum Percept Perform 40, 1332-1345, (2014).
Wood, J. N. Visual working memory for observed actions. Journal of Experimental Psychology: General 136, 639, (2007).
Wood, J. N. A core knowledge architecture of visual working memory. Journal of Experimental Psychology: Human Perception and Performance 37, 357, (2011).
Baddeley, A. D., Allen, R. J. & Hitch, G. J. Binding in visual working memory: The role of the episodic buffer. Neuropsychologia 49, 1393-1400, (2011).
Alloway, T. P., Gathercole, S. E., Willis, C. & Adams, A.-M. A structural analysis of working memory and related cognitive skills in young children. J. Exp. Child Psychol. 87, 85-106, (2004).
Gray, S. et al. The structure of working memory in young children and its relation to intelligence. Journal of Memory and Language 92, 183-201, (2017).
Shen, M., Huang, X. & Gao, Z. Object-based attention underlies the rehearsal of feature binding in visual working memory. Journal of Experimental Psychology: Human Perception and Performance 41, 479, (2015).
Matsukura, M. & Vecera, S. P. Interference between object-based attention and object-based memory. Psychonomic Bulletin & Review 16, 529-536, (2009).
Matsukura, M. & Vecera, S. P. Object-based selection from spatially-invariant representations: Evidence from a feature-report task. Attention, Perception, & Psychophysics 73, 447-457, (2011).
Logie, R. H. The functional organization and capacity limits of working memory. Current directions in Psychological science 20, 240-245, (2011).
Zhou, H. et al. A domain-general frontoparietal network interacts with domain-preferential intermediate pathways to support working memory task. Cereb. Cortex 33, 2774-2787, (2023).
King, M., Hernandez-Castillo, C. R., Poldrack, R. A., Ivry, R. B. & Diedrichsen, J. Functional boundaries in the human cerebellum revealed by a multi-domain task battery. Nat Neurosci 22, 1371-1378, (2019).
Boisgontier, M. P. Motor aging results from cerebellar neuron death. Trends Neurosci. 38, 127-128, (2015).
De Zeeuw, C. I. & Ten Brinke, M. M. Motor Learning and the Cerebellum. Cold Spring Harb Perspect Biol 7, a021683, (2015).
Itō, M. The cerebellum and neural control. (Raven press, 1984).
Holmes, G. The cerebellum of man. Brain 62, 1-30, (1939).
Ashida, R., Cerminara, N. L., Edwards, R. J., Apps, R. & Brooks, J. C. Sensorimotor, language, and working memory representation within the human cerebellum. Hum. Brain Mapp. 40, 4732-4747, (2019).
Wagner, M. J. & Luo, L. Neocortex-Cerebellum Circuits for Cognitive Processing. Trends Neurosci. 43, 42-54, (2020).
Heleven, E., van Dun, K. & Van Overwalle, F. The posterior Cerebellum is involved in constructing Social Action Sequences: An fMRI Study. Sci. Rep. 9, 11110, (2019).
Van Overwalle, F., Van de Steen, F., van Dun, K. & Heleven, E. Connectivity between the cerebrum and cerebellum during social and non-social sequencing using dynamic causal modelling. Neuroimage 206, 116326, (2020).
Leggio, M. G. et al. Cognitive sequencing impairment in patients with focal or atrophic cerebellar damage. Brain 131, 1332-1343, (2008).
Cattaneo, L. et al. Your actions in my cerebellum: subclinical deficits in action observation in patients with unilateral chronic cerebellar stroke. Cerebellum 11, 264-271, (2012).
Molinari, M. et al. Cerebellum and detection of sequences, from perception to cognition. The Cerebellum 7, 611-615, (2008).
Hasson, U., Chen, J. & Honey, C. J. Hierarchical process memory: memory as an integral component of information processing. Trends Cogn. Sci. 19, 304-313, (2015).
Xu, Y. The role of the superior intraparietal sulcus in supporting visual short-term memory for multifeature objects. J. Neurosci. 27, 11676-11686, (2007).
Parra, M. A., Della Sala, S., Logie, R. H. & Morcom, A. M. Neural correlates of shape-color binding in visual working memory. Neuropsychologia 52, 27-36, (2014).
Tseng, P., Chang, Y.-T., Chang, C.-F., Liang, W.-K. & Juan, C.-H. The critical role of phase difference in gamma oscillation within the temporoparietal network for binding visual working memory. Sci. Rep. 6, (2016).
Owen, A. M., McMillan, K. M., Laird, A. R. & Bullmore, E. N-back working memory paradigm: a meta-analysis of normative functional neuroimaging studies. Hum. Brain Mapp. 25, 46-59, (2005).
Zachariou, V. et al. Cortical iron disrupts functional connectivity networks supporting working memory performance in older adults. NeuroImage 223, 117309, (2020).
Vatansever, D. et al. Angular default mode network connectivity across working memory load. Hum. Brain Mapp. 38, 41-52, (2017).
Zuo, N., Yang, Z., Liu, Y., Li, J. & Jiang, T. Core networks and their reconfiguration patterns across cognitive loads. Hum. Brain Mapp. 39, 3546-3557, (2018).
Murphy, A. C., Bertolero, M. A., Papadopoulos, L., Lydon-Staley, D. M. & Bassett, D. S. Multimodal network dynamics underpinning working memory. Nat Commun 11, 3035, (2020).
Vanrie, J. & Verfaillie, K. Perception of biological motion: A stimulus set of human point-light actions. Behavior research methods, instruments, & computers 36, 625-629, (2004).
Shen, M., Gao, Z., Ding, X., Zhou, B. & Huang, X. Holding biological motion information in working memory. Journal of Experimental Psychology: Human Perception and Performance 40, 1332-1345, (2014).
Unsworth, N. & Spillers, G. J. Working memory capacity: Attention control, secondary memory, or both? A direct test of the dual-component model. Journal of Memory and Language 62, 392-406, (2010).
Cowan, N. The magical number 4 in short-term memory: A reconsideration of mental storage capacity. The Behavioral and brain sciences 24, 87-114, (2001).
Curby, K. M. & Gauthier, I. A visual short-term memory advantage for faces. Psychonomic bulletin & review 14, 620-628, (2007).
Gao, Z., Bentin, S. & Shen, M. Rehearsing biological motion in working memory: An EEG study. J. Cognit. Neurosci. 27, 198-209, (2015).
Muthén, B. & Muthén, L. Mplus, in Handbook of item response theory 507-518 (Chapman and Hall/CRC, 2017).
Satorra, A. & Bentler, P. M. A scaled difference chi-square test statistic for moment structure analysis. Psychometrika 66, 507-514, (2001).
DiStefano, C. Examining fit with structural equation models. (2016).
Cheung, G. W. & Rensvold, R. B. Evaluating goodness-of-fit indexes for testing measurement invariance. Structural equation modeling 9, 233-255, (2002).
Shipstead, Z., Lindsey, D. R., Marshall, R. L. & Engle, R. W. The mechanisms of working memory capacity: Primary memory, secondary memory, and attention control. Journal of Memory and Language 72, 116-141, (2014).
Cox, R. W. AFNI: Software for analysis and visualization of functional magnetic resonance neuroimages. Comput. Biomed. Res. 29, 162-173, (1996).
Yan, C. G., Wang, X. D., Zuo, X. N. & Zang, Y. F. DPABI: Data Processing & Analysis for (Resting-State) Brain Imaging. Neuroinformatics 14, 339-351, (2016).
Peng, Y. et al. MVPANI: A Toolkit With Friendly Graphical User Interface for Multivariate Pattern Analysis of Neuroimaging Data. Front Neurosci 14, 545, (2020).
Lu, X. et al. Holding Biological Motion in Working Memory: An fMRI Study. Front Hum Neurosci 10, 251, (2016).

There is NO Competing Interest.

supplementarymaterials.docx

Download PDF

Version 1

posted

You are reading this latest preprint version

Event Cache: An Independent Component in Working Memory

Status:

Version 1

Abstract

Figures

Introduction

P. W. Anderson in More is different, 1972

Results

Descriptive statistics and correlations between the WM measures

Latent variable modelling based on Baddeley and Cowan’s WM models

Event cache as an independent component in WM

Loading biological movement and non-biological movement into other factors worsens model fitting

Hierarchical structure of WM subcomponents

Differentiation of event cache from the storage of motion

Support vector regression prediction model of three WM components

The importance of different brain networks for the three WM components

The contributions of nodes to the prediction for three WM components

Virtual lesion study on the role of cerebellum in prediction of EVENT

ROI-based analysis revealed the engagement of cerebellum in event WM tasks

Discussion

Event cache as an independent psychological construct of WM

Both visual features and bindings are retained in object cache

Event cache, object cache, and CE have distinct neural correlates

Methods

Participants and experimental designs

WM tasks

Event storage tasks

Object storage tasks

Binding tasks

Central executive tasks

Other WM tasks

Behavioral analysis

MRI imaging parameters and pre-processing

Brain-behavior prediction analyses

ROI analyses of event WM tasks

Declarations

Acknowledgment

References

Additional Declarations

Supplementary Files

Status:

Version 1