Task‐specific topology of brain networks supporting working memory and inhibition

Abstract

Network neuroscience explores the brain's connectome, demonstrating that dynamic neural networks support cognitive functions. This study investigates how distinct cognitive abilities—working memory and cognitive inhibitory control—are supported by unique brain network configurations constructed by estimating whole‐brain networks using mutual information. The study involved 195 participants who completed the Sternberg Item Recognition task and Flanker tasks while undergoing electroencephalography recording. A mixed‐effects linear model analyzed the influence of network metrics on cognitive performance, considering individual differences and task‐specific dynamics. The findings indicate that working memory and cognitive inhibitory control are associated with different network attributes, with working memory relying on distributed networks and cognitive inhibitory control on more segregated ones. Our analysis suggests that both strong and weak connections contribute to cognitive processes, with weak connections potentially leading to a more stable and support networks of memory and cognitive inhibitory control. The findings indirectly support the network neuroscience theory of intelligence, suggesting different functional topology of networks inherent to various cognitive functions. Nevertheless, we propose that understanding individual variations in cognitive abilities requires recognizing both shared and unique processes within the brain's network dynamics.

We used mixed‐effects linear regression to predict accuracy and response time in Flanker and Sternberg tasks based on measures of network topology during tasks and resting state. Our results suggest that the two processes have unique topological features in different frequency bands but also share common processes.

Practitioner Points.

1. We estimate network features of resting‐state and task‐induced electroencephalography networks

2. Working memory mostly relies on a distributed network, while cognitive inhibitory control depends on a more segregated one.

3. Individual differences in cognitive functions might be related to both shared and unique processes of brain dynamics.

1. INTRODUCTION

Network neuroscience combines graph theory and neuroimaging to decipher the brain's intricate web of connections, illuminating how neural networks underpin cognitive functions by viewing the brain as a system of interlinked regions (Deco & Kringelbach, 2017; Sporns & Betzel, 2016). According to research, the brain functions as a small‐world network with high clustering and short path lengths, which allow for both specialized (segregated) and holistic (integrated) processing (Cohen & D'Esposito, 2016; Deco et al., 2015). The brain's dynamic network adapts to the cognitive demands of tasks like memory retrieval or problem‐solving by reconfiguring connections to optimize function (Braun et al., 2015; Gonzalez‐Castillo et al., 2015). Individual differences in network structure and adaptability are thought to contribute to cognitive performance variations, emphasizing the importance of the brain's dynamic reconfiguration in response to cognitive challenges (Cole et al., 2014; Finn et al., 2015; Krienen et al., 2014).

The investigation of the brain mechanisms of cognition must take into account the heterogeneity nature of cognitive functions. Theoretical models of intelligence, such as the Cattell–Horn–Carroll theory, have been instrumental in conceptualizing intelligence as a hierarchical structure of abilities, ranging from general intelligence to broad and narrow cognitive abilities (Carroll, 1993; McGrew, 2009). However, the emergence of network neuroscience has provided a new perspective on cognition, linking these hierarchical models to the brain's network architecture. The Network neuroscience theory of intelligence (Barbey, 2018) posits that intelligence stems from the dynamic interplay of the brain's networks, with the adaptability and flexibility of these networks being crucial for high‐level cognitive functioning. According to this theory, different levels of intelligence correspond to the brain's ability to reconfigure its networks. The theory further suggests that the manifestation of crystallized and fluid intelligence can be traced to the nature of connections within these networks—strong, stable connections and easy‐to‐reach brain states facilitate verbal intelligence, while weaker, more variable connections and hard‐to‐reach states might underpin nonverbal intelligence. The theory is consistent with findings that associate efficient network reconfiguration with fluid intelligence, suggesting that the brain's ability to switch between different network states in response to cognitive demands is a basis of cognitive ability (Anderson & Barbey, 2023; Bassett et al., 2011; Dubois et al., 2018; Wen et al., 2015). However, several recent research have discovered the reverse connections between brain reconfiguration and intelligence: more stability of brain networks was connected to higher intelligence (Hilger et al., 2017, 2020; Pamplona et al., 2015; Thiele et al., 2022). Other research were unable to establish a connection between general factor of intelligence on functional magnetic resonance imaging (fMRI) data of the Human Connectome Project (HCP; Kruschwitz et al., 2018) and replicate the association between general intelligence and network measures on four independent datasets (Metzen et al., 2024).

In this study, we aimed to expand on the premise that various types of intelligence are supported by distinct network bases, by analyzing two cognitive abilities closely related to intelligence and general cognitive functioning: working memory (WM) and cognitive inhibitory control.

WM is the capacity to keep and modify important information temporarily for cognitive tasks, playing a vital role in reasoning and the guiding of decision‐making and behavior (Barak & Tsodyks, 2014; Miller et al., 2018). From empirical to behavioral brain investigations, a strong relationship has been revealed between WM and general intelligence (g). In particular, meta‐analyses have shown that the average correlation between WM and g measures was 0.479 (Ackerman et al., 2005), and WM capacity and fluid intelligence shared about 50% of their variance (Kane et al., 2005).

Cognitive inhibitory control (Diamond, 2013) often known as inhibition, is the ability to suppress irrelevant or distracting stimuli, allowing for appropriate responses in the context of a given task (Rondeel et al., 2015). Cognitive inhibitory control, like WM, is essential for both ordinary living and higher cognitive activity (Aron, 2007). Conversely, the relationship between intelligence and cognitive inhibitory control is more contentious, with studies yielding mixed results across various ages (Duan & Shi, 2011; Friedman et al., 2006).

These two processes—WM and cognitive control—are deeply interconnected from a psychological standpoint, both contributing significantly to overall cognitive functioning. In particular, Engle and Kane, in their two‐factor theory of cognitive control, argued for the separation of two functions that form executive attention: goal maintenance and resolution of response conflict. While emphasizing the interaction between them within the framework of research on individual differences in WM capacity (Engle & Kane, 2004). Braver et al. identify the several internal mechanisms of control processes in WM, including information selection, storage and updating it over an appropriate time period, protecting this information from interference, and using this information to structure and bias other cognitive systems (Braver et al., 2008).

Overall, WM and cognitive inhibitory control are central to cognitive functions in general and intelligence in particular. These two cognitive functions are closely connected, with cognitive inhibitory control playing an important role in WM, assisting in the retention of information in the memory.

From the perspective of brain mechanisms of cognition, evidence for the association between WM and cognitive inhibitory control with various aspects of brain dynamics is more robust than the associations with general intelligence.

Recent fMRI studies have identified that WM representations are maintained in a rather distributed network, including the prefrontal, parietal, and sensory cortices, with the posterior intraparietal sulcus playing a key role in WM capacity as a central hub in WM‐related networks (Sreenivasan, Curtis, & D'Esposito, 2014; Sreenivasan, Gratton, et al., 2014; Sreenivasan, Vytlacil, & D'Esposito, 2014; Yamashita et al., 2018).

Electrophysiological data‐driven investigations associated WM with engagement of the prefrontal cortex (PFC), temporal, and parietal lobes involvement (Avery et al., 2020; Finc et al., 2020; Murphy et al., 2020; Sala‐Llonch et al., 2012). Those areas are involved in extensive networks, such as the frontoparietal network (FPN; Emch et al., 2019) and the default mode network (S. Liu, Poh, et al., 2018; Murphy et al., 2020). The FPN could be divided into two distinct subnetworks: one aligned with the default‐mode network and the other with the dorsal attention system with competition between them influencing the WM performance (Murphy et al., 2020).

The WM neural underpinnings exhibit pronounced frequency‐related specificity. The theta bands are associated with control mechanisms and executive functions (Sauseng & Klimesch, 2008), while alpha, beta, and gamma bands are associated with to encoding and maintenance (Roux et al., 2012; Wianda & Ross, 2019), although there is some controversy regarding the effects of higher‐frequency bands (Pavlov & Kotchoubey, 2022).

The WM neural foundation appear to be influenced by the nature of the task at hand, reflecting the demands, specific to the type of information that needs to be processed and potentially indicating high demands in the configuration and reconfiguration of the networks responsible for the current task; for instance, verbal WM tasks typically activate the Broca's area, while spatial WM tasks draw on a more extensive network of brain regions (Bernal et al., 2015; Ren et al., 2019; Rogalsky et al., 2008).

Studies have been able to distinguish between several phases of WM and their respective tasks, namely memorization (encoding), maintenance (or storage), and retrieval (recollection), which were found to induce changes in brain activation and brain network topology, possibly due to the specifics of information processing in each particular phase.

It was discovered that the encoding phase exhibited a small‐world topology across all frequency bands, indicating that optimal global and local information processing and network configuration were connected with task performance. The small‐world topology persisted but decreased during maintenance and rehearsal phases. Furthermore, information flow between brain regions varied across WM phases, reflecting the different roles of brain areas during different phases of information processing (Toppi et al., 2018). Other studies demonstrated a more sophisticated and frequency‐dependent pattern of topological features (Ernston & Adamovich, 2023) and temporal, spatial, and spectral dynamics related to WM phases and WM components, such as central executive component and storage units (Rossi et al., 2023). As such, in the alpha band, networks show increased integration and centrality with reduced segregation in early phases, which decreases in later phases. Conversely, lower beta rhythm networks exhibit a significant increase in integration from the storage to the retrieval phase.

Cognitive inhibitory control has been demonstrated to involve neural networks and cognitive functions at various levels (Herstel & Wierenga, 2021). Typically, it is linked to the activity within the cognitive control network (CCN), which includes the functioning of the prefrontal cortex, the dorsal anterior cingulate cortex, and the lateral prefrontal cortex (Cole & Schneider, 2007). Several studies suggested that the reconfiguration of CCN is crucial for the effective filtering of information (Bartholomew et al., 2019; Cocchi et al., 2011; Marek et al., 2015; Menon & D'Esposito, 2022; Spielberg et al., 2015). Specifically, the theta band FPN was active during conflicting interference, and the conflict effect on graph measures diminished over time (J. Liu et al., 2024).

fMRI studies on the Eriksen flanker task have identified several brain regions involved in task performance, including the lateral prefrontal cortices, supplementary motor area, superior parietal lobe, and anterior cingulate cortex (Iannaccone et al., 2015; Siemann et al., 2016) These studies often emphasize the anterior cingulate's role in conflict monitoring, whereas frontal and parietal regions contributing to visual selective attention are less frequently highlighted. Similarly, electroencephalography (EEG) studies have corroborated the role of the anterior cingulate, noting that increased conflict during the task leads to greater frontal and parietal midline theta activity (Haciahmet et al., 2021). Cognitive inhibitory control is also linked to fluctuations in alpha band power, characterized by increases in task‐irrelevant regions and decreases in task‐relevant regions (McDermott et al., 2017; Suzuki et al., 2018) and demonstrating task‐related specificity (Xie et al., 2017).

To summarize, both WM and cognitive inhibitory control are connected to numerous neural mechanisms involved in information manipulation within these cognitive processes. These processes overlap in the activity of the prefrontal and parietal cortices, suggesting the existence of a “core” functional architecture that supports cognition (Krienen et al., 2014). The key mechanism here is the inclusion of the prefrontal cortex (Menon & D'Esposito, 2022).

The PFC is essential for active goal maintenance and using goals to influence ongoing processing, demonstrating its central role in both WM and cognitive inhibitory control. Networks involving the PFC, like the FPN, mediate different types of cognitive performance, specializing the information processing and ignoring distracting stimuli in WM, effectively removing them from focus (Buschman & Miller, 2007; Miller et al., 2018). The prefrontal cortex itself does not exhibit a distinct activation pattern for WM, but is responsible for resource allocation, attention focusing, and modulating the posterior regions, primarily the parietal and temporal cortices, where information about the stimuli is encoded (Riley & Constantinidis, 2015).

Miller et al. (2018) also explain the frequency‐related separation of information processing in WM by suggesting that low‐frequency activity suppresses high‐frequency bursts, thereby regulating information maintenance in WM.

Another aspect of the study is related to the individual differences in brain foundation of cognitive functions. Studies often focus on group‐level findings, although there is an increasing acknowledgement of the significance of individual differences in brain network measures. Studies have found that the degree of similarity between an individual's brain connectivity in resting and active states can correlate with behavioral patterns, suggesting a personalized basis for cognition (Tompson et al., 2018). This individual‐specific manifestation of cognition may be influenced by a person's unique cognitive styles and traits. Several studies (Bansal et al., 2019; S. Liu, Poh, et al., 2018) observed a substantial interindividual variability in functional connectivity. For example, the FPN and cingulo‐opercular networks exhibited strong individual variability, suggesting that cognitive control systems show greater between‐subject variability compared to sensory processing systems (Bansal et al., 2019). The structure of WM networks showed significant interindividual heterogeneity (S. Liu, Poh, et al., 2018). The study of J. Liu, Liao, et al. (2018) found that the largest variations in dynamic functional connectivity (DFC) across individuals were observed in both higher‐order systems (such as the default mode, dorsal attention, and fronto‐parietal systems) and primary systems (such as visual and sensorimotor systems). However, higher‐order systems contributed significantly more to individual identification and cognition prediction than primary systems. This suggests that DFC patterns in higher‐order systems carry more unique information reflecting individual cognitive abilities or demographic characteristics compared to primary systems. This divergence may be due to developmental and functional differences: the extended developmental period and cognitive processing complexity of higher‐order systems likely enhance the impact of variable environmental factors during individual development, resulting in increased interindividual variability, rather than early‐developed primary systems, that focus on basic visual and sensorimotor processing. These studies underscore the importance of thorough investigation of individual variability in the topology of higher‐order brain networks associated with cognition. Since both WM and cognitive inhibitory control are supported by networks with significant individual variability, accounting for this variability is crucial for accurately estimating the network topology associated with cognitive functions, as failure to account for interindividual variability may result in spurious results.

A layer of complexity in functional connectivity analysis arises from the mental state of the participant induced by the given task. Two main conditions are commonly used: the resting state and cognitive performance or task engagement. The resting state typically refers to a period without any specific mental activity, and studies in this condition are predominantly focused on assessing the stable internal architecture of networks (Smith et al., 2009). Task‐related conditions are associated with the detection of activity that supports an ongoing task. A comparison of functional connectivity in the resting state and during various tasks reveals considerable changes (Spadone et al., 2015). According to data from the HCP (Barch et al., 2013), 38–76% of connections in functional connectivity networks are different across the two states (Kelso, 2012). Nonetheless, the authors note a high similarity in functional networks in both states and individual differences observed in the resting state persist during tasks (Shah et al., 2016) and these individual differences vary with cognitive state (Geerligs et al., 2015). Individual differences in functional connectivity in both resting state and task‐related conditions contribute to the identification of specific individuals (Salehi et al., 2019) and mental states (Shine & Poldrack, 2018; Suhail et al., 2022). This specificity of individual networks also extends to transitions between states, such as from the resting state to cognitive task conditions (Braun et al., 2015; Gonzalez‐Castillo et al., 2015).

Overall, the objective of this article is to identify common and unique topological features of functional connectivity that are related to cognitive performance across two cognitive functions: WM and cognitive inhibitory control. We hypothesize that distinct cognitive processes rely on different brain network configurations as various types of information manipulation, specifically memorization and retrieval in WM and inhibition of distracting information, are supported by specific topological features. Furthermore, due to the high interrelatedness of cognitive processes, there will be an overlap in their topological properties, representing a core that supports general cognition.

2. MATERIALS AND METHODS

2.1. Participants

The participants were recruited through announcements in universities. Participation was voluntary and without any reward or remuneration. Then, 195 participants took part in the study. The age ranged from 17 to 39 years (M = 20.3, SD = 3.35, and 37% identified as female). Participants had no recorded history of psychiatric or neurological disorders, or head traumas. All participants gave their consent to participate in the study. All procedures were approved by the Ethics Committee of the Psychological Institute of the Russian Academy of Education (protocol code 2019/2‐10, date of approval 11.02.2019).

2.2. Experimental procedure

The study involved two cognitive tasks: the Sternberg Item Recognition Task and the Flanker Task, which were administered in an alternating sequence across the sample. Each recording session started with a resting state. During resting state, participants were asked to relax, not to focus on anything specific, and to stay awake for 10 min. They received spoken commands every 2 min to either open or close their eyes. We only analyzed the EEG data collected while their eyes were open. Each cognitive task took from 15 to 20 min, so the overall duration for EEG acquisition was around 45–50 min. A state during which an EEG was recorded, whether it represents a resting state or cognitive tasks, is referred to as a “condition” in the manuscript.

2.3. Cognitive tasks

Sternberg Item Recognition Paradigm or Sternberg Task is a delayed stimulus identification task, which allows assessing WM performance when solving problems of varying complexity. The task also allows distinguishing between such indicators as “search speed when performing a task” (mental scanning speed) and the direct success of completing a WM task.

The task consists of 13 training and 72 test trials. In each trial, subjects are presented with a fixation cross of 0.5 s in duration followed by a pseudo‐random set of numbers that they need to memorize. The set length varies from one to six digits (one to three digits in train trials and three to six digits in test trials with equal number of trials per digit), each of which is presented for 1.2 s. This is followed by a 2 s pause, and then a check digit. Subjects must press the right arrow button on the keyboard to answer “Yes, this is one of the memorized digits” or the left arrow button “No, this is a new digit” (correct and incorrect control digits are equally likely to appear), after which the control stimulus disappears. During training trials, feedback was given in the form of text: “Correct” or “Error.” The task has the following parameters: number of correct and incorrect answers, and reaction time for each trial. For analysis, Sternberg performance as number of correct answers (accuracy) and Sternberg RT as reaction time were used.

Eriksen Flanker Task is a set of response inhibition tasks that assesses cognitive inhibition, specifically the ability to inhibit responses that are inappropriate for a given situational context. The participant was presented with both targeted and distracting stimuli, specifically, an arrow pointing right or left, surrounded by other arrows. Congruent condition in the task—the target arrow is surrounded by other arrows pointing in the same direction. Incongruent condition—additional arrows point in the opposite direction. The participant's task is to indicate in which direction the target arrow is pointing. The fixation cross was presented at the beginning of each trial, followed by the arrows after 0.5 s. The participant had 1 s to respond, and upon pressing a key or at the end of the response time (RT), the arrows disappeared, initiating a 1‐s intertrial interval.

A total of 200 test trials were presented, with half of them congruent and the other half being incongruent. As part of the assessment of cognitive inhibition, only tasks for incongruent condition were used as they impose higher load on the WM (Xie et al., 2017). Number of correct answers (accuracy) and reaction time were used in the study. Examples of stimuli for both tasks are presented in Appendix 3.

2.4. Preprocessing of EEG data

The EEG data were recorded using a 64‐electrode cap, set up according to the 10–10 international system, with a Brain Products ActiChamp amplifier. EEG was recorded in a sound‐attenuated and electrically shielded room, with dim lighting. High conductive chloride gel was used to keep the impedance of the electrodes below 25 kOhm.

The EEG data were recorded in real‐time using the Brain Products PyCorder system, which recorded continuously at a 500 Hz sample rate without initially applying any filters. The reference point for the recording was the Cz electrode. During EEG‐data preprocessing, we switched to a REST reference (Yao, 2001) and reduced the sample rate to 256 Hz. Next, a high‐pass filter at 1 Hz was applied. To clean the data, we used ICA decomposition to identify and remove typical interference like eye blinks, eye movements, and muscle artifacts. IClabel (Pion‐Tonachini et al., 2019) was used to classify the components into brain and non‐brain activity. Components marked as non‐brain activity were removed from the data. On average, 22.7 ± 5.9 components (~36 ± 9%) were marked as artifactual and eliminated from the data.

The EEG data were divided into 1‐s consequent nonoverlapping segments for the resting state and 800 ms segments for the Flanker task and the Sternberg task Retrieval Phase, and 1‐s segments for the Sternberg task Memorization Phase. The Sternberg task Memorization Phase started with the presentation of the set, and the Sternberg task Retrieval Phase started with the presentation of the target number.

Segments containing artifacts were rejected using autoreject algorithm (Jas et al., 2017). The data were considered fit for the study as long as less than 20% of epochs were marked as bad across all the conditions.

2.5. Estimation of functional connectivity and network topology

The investigation was performed in four distinct frequency bands: theta (4–8 Hz), alpha (8–13 Hz), low beta (13–20 Hz), and high beta (20–30 Hz). The exploration of inter‐frequency interactions was beyond the scope of this study. Findings for each experimental condition across these specified frequency bands were presented.

In this study, we sought to account for a possible spectrum of individual brain states associated with cognitive performance while acknowledging the dynamic nature of brain activity and accounting for interindividual differences in the sample. To do this, we divide each condition into five nonoverlapping intervals, each comprising 20% of the relevant stimuli in the task, beginning with the first stimuli. The number of intervals was determined to consistently contain at least 10 stimuli in each interval for both tasks, with the possibility of correcting for some outliers through averaging. Within each interval the Gaussian‐Copula mutual‐information estimator (Ince et al., 2017) was used to calculate the functional connectivity for each stimulus that was then averaged within each interval to obtain the averaged connectivity matrix for the interval. Total of five connectivity matrices were constructed for each condition. Each connectivity matrix is a representation of pairwise connections between EEG electrodes, forming an adjacency matrix of a complete graph that includes all possible connections between the nodes of the graph. To investigate the potential effects of spatial leakage on our estimates of mutual information, we compared mutual information for neighboring and non‐neighboring electrodes with differences between them being considered insignificant. Details of this test can be found in Appendix 1.

Adjacency matrices were transformed into networks through quartile thresholding (Adamovich et al., 2022; van den Heuvel et al., 2017). This process attempted to remove any weak or spurious connections. Since the choice of a threshold value for removing spurious graph connections is arbitrary and there are no established guidelines for choosing this value, it is prudent to conduct an analysis for several thresholds. The threshold of 0.5 was chosen because our previous research (Zakharov et al., 2020) demonstrated that connectivity values obtained at this threshold are associated with cognitive characteristics. Each threshold represents a proportion of weakest connections from the general distribution that would be removed from the network. Entries in each connectivity matrix falling below the respective quantile threshold were reassigned a value of zero. To preserve network integrity and avoid isolated nodes, any nodes that became disconnected post‐thresholding were reconnected to their two strongest links. The Leiden algorithm was used to identify modules in the networks (Traag et al., 2019). Each module (or community) is a group of nodes that are more densely connected to each other than to nodes in other groups.

Network structure was characterized using several metrics, including:

1. Average path length. The average path length is a measure of the average number of steps (or cost of transfer between nodes) along the shortest paths for all possible pairs of network nodes. In this study, higher path length indicates higher cost of transfer.

2. Betweenness centrality: This quantifies the number of times a node acts as a bridge along the shortest path between two other nodes. Nodes with high betweenness centrality play a critical role in the flow of information within the network. Eigenvector centrality, this measures a node's influence by considering the number and quality of its connections; nodes with high eigenvector centrality are connected to many nodes that have high centrality.

3. Clustering coefficient: This is a measure of the degree to which nodes in a network tend to cluster together. A higher clustering coefficient indicates a greater “local” connectivity, which in EEG studies can suggest local processing or functional segregation within the brain. It reflects the presence of tightly connected groups of regions that may specialize in specific cognitive functions.

4. Modularity: This is a measure of the strength of division of a network into modules. Networks with high modularity have dense connections between the nodes within modules but sparse connections between nodes in different modules.

5. Participation coefficient: This quantifies how well a node interacts with multiple modules in the network. A high participation coefficient indicates that a node has several connections to nodes in different modules.

6. Rich club coefficient: This measures the tendency of high‐degree nodes (hubs) to be more interconnected than expected by chance. If a network has a high rich club coefficient, it suggests the presence of a “rich club”—a cohesive group of central nodes that are highly interlinked and may coordinate complex network functions.

2.6. Statistical analysis of behavioral measures

Previous research has shown that information processing speed (RT) is an important predictor of cognitive performance, along with the actual accuracy of the response (Meiran & Shahar, 2018; Schubert et al., 2017), so cognitive performance in the two tasks was quantified through accuracy and RT, yielding four behavioral measures in total. The approach for estimating behavioral measures mirrored the connectivity analysis; each task was divided into five intervals for each participant, with average accuracy and RT calculated separately for each interval, resulting in five data points per cognitive measure that corresponded to the connectivity matrices. For the Flanker task, data were exclusively drawn from incongruent trials. To achieve a more Gaussian distribution of RTs, we applied a Box‐Cox transformation.

Spearman correlation was used to examine the structure and interrelations between the four behavioral measures. To evaluate the potential effect of fatigue and ensure that performance is consistent across the task and does not decline over time, a one‐way repeated measures ANOVA with eta‐squared as effect size was conducted to estimate the differences between the five intervals.

2.7. Statistical analysis of brain–behavior relationships

We employed a mixed‐effects linear model (Baayen et al., 2008) with random intercepts to estimate the impact of all network metrics on cognitive outcomes, treating participant ID as a random effect and conducting the analysis for each frequency band separately. Prior to fitting the model, all variables were standardized. To improve the robustness of our findings, a bootstrap resampling method was implemented. Across 2000 iterations, the model was fit to a bootstrapped dataset, assembled by randomly sampling participants with replacement, maintaining the original dataset's size.

The analysis is based on the consequent distribution of linear regression prediction coefficients. A regression effect was considered significant if the two‐tailed 10% confidence interval for the distribution did not include zero.

Because of the previously observed state‐specificity of brain networks, brain–behavior interactions were examined separately for the resting state, cognitive tasks, and shifts between two states. Furthermore, we determined the topological features of resting‐state networks and their relationship with cognitive task performance, assessing the influence of the brain's background information exchange processes on performance under cognitive load. We also explored differences between these states to determine how the reconfiguration between a resting state and a state of cognitive task affects cognitive performance. To calculate the reconfiguration of networks between the resting state and the cognitive load state, the resting state value was subtracted from the corresponding metric obtained during the cognitive load. An increase in values along the scale indicates that the metric value of the networks is higher within the task with cognitive load.

2.8. Software

We used MNE‐python (Gramfort et al., 2014) for EEG preprocessing, the Frites software package (Combrisson et al., 2022) for estimation of functional connectivity, networkX (Hagberg et al., 2008) for graph measures calculation and lme4 (Bates et al., 2015) for modeling.

3. RESULTS

The results section is structured to first describe the interrelations between behavioral measures obtained in the study, we then investigate the effects of network topology on behavioral measures in four conditions, namely: the most robust network characteristics estimated during the task that are associated with task performance. The brain's intrinsic characteristics of information exchange as predictors of performance in cognitive tasks, based on resting EEG data are evaluated. Following this, the reconfiguration between resting‐state and task‐induced functional connectivity is assessed as a factor influencing performance in cognitive tasks. Finally, the effect of functional connectivity at lower thresholds is analyzed.

3.1. Analysis of behavioral measures

Table 1 displays descriptive statistics for the five segments and overall cognitive tasks, along with comparison results. One‐way repeated measures ANOVA indicates the differences in the behavioral score related to poorer performance in the first 20% of the Flanker accuracy and RT in both tasks, throughout the rest of the task, performance was consistent.

TABLE 1.

Descriptive statistics for whole cognitive tasks and five segments and results of repeated‐measures ANOVA.

Measure	Statistics	Task segment						RM‐ANOVA	Eta‐squared
		1	2	3	4	5	Full‐length
Sternberg accuracy	Mean (SD)	0.88 (0.14)	0.91 (0.13)	0.90 (0.11)	0.90 (0.12)	0.90 (0.12)	0.90 (0.12)	F(4, 408) = 1.68	0.016
	Median (IQR)	0.92 (0.14)	0.93 (0.14)	0.93 (0.15)	0.92 (0.15)	0.94 (0.13)	0.93 (0.14)
Sternberg RT	Mean (SD)	1.03 (0.43)	0.89 (0.23)	0.90 (0.42)	0.92 (0.48)	0.87 (0.27)	0.92 (0.38)	F(4, 408) = 11.3**	0.1
	Median (IQR)	0.93 (0.32)	0.86 (0.30)	0.81 (0.28)	0.83 (0.27)	0.81 (0.22)	0.85 (0.30)
Flanker accuracy	Mean (SD)	0.28 (0.22)	0.37 (0.28)	0.37 (0.28)	0.36 (0.27)	0.39 (0.27)	0.36 (0.27)	F(4, 408) = 2.583*	0.025
	Median (IQR)	0.25 (0.35)	0.35 (0.50)	0.35 (0.45)	0.35 (0.45)	0.40 (0.45)	0.35 (0.45)
Flanker RT	Mean (SD)	0.73 (0.25)	0.80 (0.24)	0.82 (0.22)	0.81 (0.23)	0.83 (0.20)	0.79 (0.23)	F(4, 408) = 17.4**	0.168
	Median (IQR)	0.85 (0.47)	0.90 (0.13)	0.92 (0.11)	0.92 (0.17)	0.91 (0.07)	0.90 (0.19)

^* p < .05,

^** p < .01.

Correlations were obtained between accuracy and RT in the Sternberg task. The average RT demonstrates a small negative correlation with task accuracy (Spearman's Rho = −0.098, p < .05), which may indicate that the task was not very difficult for the respondents and they did not reach their full WM capacity. On the other hand, the positive correlation of the Flanker task's RT with its accuracy (Spearman's Rho = 0.199, p < .001) corresponds with the results expected in the conflict paradigm in inhibition tasks (Hübner & Töbel, 2019). Correlations between the performance measures of the Sternberg task and the Flanker task indicate a connection between WM and inhibition, suggesting an important role of WM in conflict situations (Pratt et al., 2011). The results of correlational analysis are presented in Table 2.

TABLE 2.

Correlation between cognitive measures.

	Sternberg accuracy	Sternberg RT	Flanker accuracy	Flanker RT
Sternberg accuracy	—
Sternberg RT	−0.098*	—
Flanker accuracy	0.052	−0.240***	—
Flanker RT	0.073	−0.021*	0.199***	—

Note: RTs were box‐cox transformed.

^* p < .05,

^*** p < .001.

3.2. The effect of task‐induced functional connectivity on performance in cognitive tasks

The aim of this section is to describe the most robust network characteristics estimated during the task associated with the task performance. The significant results reported in this section are presented in Figure 1 and are marked in color. A complete table with descriptive statistics of the resulting distributions is available in Appendix 2 under the sheet titled “Threshold 0.8.”

FIGURE 1.

The significant effects of network measures during cognitive tasks on behavioral measures. Within each table, graph topology measures are arranged row‐wise, and frequency bands are visualized column‐wise. The mean regression coefficient is reported with the standard deviation in parentheses. Significant effects are marked with color: Positive effects in orange and negative effects in light green. A more pronounced effect is associated with a more saturated color.

The relationship between path length and execution time (inversely) in alpha and theta bands is the most significant. In the alpha band, path length during the memorization phase is positively associated with task execution accuracy. A similar relationship is observed in the high beta band, suggesting that increased information exchange between different brain areas leads to decreased RT and improved task performance. Betweenness centrality is predominantly associated with RT in low and high beta bands, both during the memorization and retrieval phases with reduced centrality levels linked to increased RTs. Higher clustering coefficient during the memorization phase in the high beta band is associated with increased accuracy while in the alpha band it is associated with increased RT. Eigenvector centrality during memorization has a negative relation to RT in the alpha band. Modularity has clear negative relationships with RT during the memorization phase across all bands, and furthermore with retrieval phase in the low beta band. Increased modularity in the high beta band for both phases is associated with faster and more accurate response. In the alpha and low beta bands, an increase in the participation index is related to more successful task performance, while a decrease in theta band is associated with better performance. A lower rich club coefficient in theta band leads to longer execution time, while an increase in the memorization phase is associated with a better WM performance.

There were no significant associations between the Flanker task and path length. For the Flanker task, lower average path lengths are associated with higher RT. Betweenness centrality is inversely associated with RT, with higher betweenness in the alpha band and lower betweenness in the lower beta band corresponding to higher RT. Participants with higher clustering coefficient values in the alpha band and theta band have lower accuracy and slower RT. An increase in eigenvector centrality value in the low beta band is associated with higher performance accuracy and execution time. Similar associations are also observed in high beta bands but only with execution time. An increase in modularity in the theta rhythm is associated with more successful task performance. A decrease in the participation index in the alpha band is associated with reduced accuracy and increased reaction time. The rich‐club index is associated with task execution accuracy only in the alpha band.

3.3. The effect of resting‐state functional connectivity on performance in cognitive tasks

The second section focuses on the evaluation of the brain's intrinsic characteristics of information exchange as predictors of performance in cognitive tasks, based on resting EEG data. The significant results reported in this section are presented in Figure 2 and are marked in color. A complete table with descriptive statistics of the resulting distributions is available in Appendix 2 under the sheet titled “Threshold 0.8.”

FIGURE 2.

The significant effects of network measures during resting state on behavioral measures. Within each table, graph topology measures are arranged row‐wise, and frequency bands are visualized column‐wise. The mean regression coefficient is reported with the standard deviation in parentheses. Significant effects are marked with color: Positive effects in orange and negative effects in light green. A more pronounced effect is associated with a more saturated color.

It has been found that the path length between brain regions does not significantly relate to the performance in Sternberg and Flanker tasks. The betweenness centrality coefficient in alpha band shows a unique relationship with the accuracy in the Flanker task. The clustering coefficient demonstrates a connection with performance in both tasks: for the Flanker task, it relates with the accuracy in both beta bands and the RT positively in the alpha band and negatively in the low beta band, while in the Sternberg task, it is inversely linked with behavioral measures: positively with the accuracy and negatively with RT Eigenvector centrality is another significant predictor of performance across tasks; in the Flanker task, it negatively relates to the RT in theta band and alpha band, and to the accuracy in alpha band and both beta bands. In the Sternberg task, higher RT is associated with lower eigenvector centrality in the alpha band and high beta band. Modularity negatively corresponds to accuracy in the Sternberg task in theta band and to both accuracy and RT in the Flanker task within the alpha band. Furthermore, the lower participation index is tied to the higher accuracy of the Flanker task in the low beta band. Finally, the RT for the Sternberg task at various frequencies (theta, alpha, and both beta bands) is negatively linked with the brain's rich club network, indicating that nodes with more extensive connections to other nodes in the network are associated with longer performance times.

3.4. Effects of reconfiguration between resting‐state and task‐induced functional connectivity on performance in cognitive tasks

The purpose of this section is to assess the reconfiguration between the two states as a factor of performance in cognitive tasks. The significant results reported in this section are presented Figure 3 and are marked in color. A complete table with descriptive statistics of the resulting distributions is available in Appendix 2 under the sheet titled “Threshold 0.8.”

FIGURE 3.

The significant effects of reconfiguration between resting state and cognitive tasks on behavioral measures. Within each table, graph topology measures are arranged row‐wise, and frequency bands are visualized column‐wise. The mean regression coefficient is reported with the standard deviation in parentheses. Significant effects are marked with color: Positive effects in orange and negative effects in light green. A more pronounced effect is associated with a more saturated color.

For the Sternberg task, neither path length nor betweenness centrality are related to task accuracy. The clustering coefficient in the theta band is negatively associated with accuracy during the memorization phase. In the alpha band and both beta bands, it is similarly associated with the retrieval phase. Clustering coefficient is a sole predictor of RT with positive associations in alpha band and lower beta band. Eigenvector centrality during retrieval positively affects accuracy in the theta band but has a negative effect during memorization in the alpha band. Accuracy is related to measures in theta band: positively with modularity during memorization phase and negatively with participation index during memory retrieval. In low beta band, lower values of rich club coefficient during retrieval are associated with higher task accuracy.

For the Flanker task, path length is negatively related with RT across all frequency bands (theta, alpha, low, and high beta), indicating that greater reconfiguration toward enhancing information exchange between the resting state and the Flanker task is associated with longer reaction times. Centrality has varying effects on RT at different frequency bands: positively in the alpha band and negatively in the low beta band. Clustering coefficient negatively affects accuracy in the theta band and RT in the alpha band, suggesting that a higher level of clustering is associated with lower correctness and longer RTs. Eigenvector centrality positively relates to accuracy in alpha, low, and high beta bands, suggesting that a higher level of eigenvector centrality compared to resting state at these frequencies is linked to higher task correctness. Modularity has a positive impact on task accuracy in the theta band and on RT in the alpha band, suggesting that a higher level of modularity compared to resting state is associated with higher task correctness. Participation coefficient does not demonstrate significant relationships with performance in the Flanker task. The rich club is positively associated with task accuracy in alpha and low beta bands in the Flanker task, with the higher level of rich clubs linked to higher task accuracy.

3.5. Effect of functional connectivity at lower thresholds

In this section, we examine the additional association between brain network topology and performance in the Sternberg task and the Flanker task by incorporating weaker connections in the network. The results reported in this section are presented in Figures 4, 5, and 6. A complete table with descriptive statistics of the resulting distributions is available in Appendix 2 under the sheet titled “Threshold 0.5.”

FIGURE 4.

The significant effects of network measures during cognitive tasks on behavioral measures at 0.5 threshold. Within each table, graph topology measures are arranged row‐wise, and frequency bands are visualized column‐wise. The mean regression coefficient is reported with the standard deviation in parentheses. Significant effects are marked with color: Positive effects in orange and negative effects in light green. A more pronounced effect is associated with a more saturated color.

FIGURE 5.

The significant effects of network measures during resting state on behavioral measures at 0.5 threshold. Within each table, graph topology measures are arranged row‐wise, and frequency bands are visualized column‐wise. The mean regression coefficient is reported with the standard deviation in parentheses. Significant effects are marked with color: Positive effects in orange and negative effects in light green. A more pronounced effect is associated with a more saturated color.

FIGURE 6.

The significant effects of reconfiguration between resting state and cognitive tasks on behavioral measures at threshold 0.5. Within each table, graph topology measures are arranged row‐wise, and frequency bands are visualized column‐wise. The mean regression coefficient is reported with the standard deviation in parentheses. Significant effects are marked with color: Positive effects in orange and negative effects in light green. A more pronounced effect is associated with a more saturated color.

The comparison of thresholds (please see Figure 4) demonstrates that incorporating weaker connections in the network results in extra linkages with cognitive traits. During task execution, this effect is particularly noticeable in the Sternberg task and the Flanker task's reaction time. The Flanker task accuracy shows individual additional associations. In the Sternberg task, enhanced accuracy is linked to a reduced clustering coefficient across frequencies during the memorization phase, indicating less local interconnectivity. Furthermore, increased modularity and rich club coefficients in the theta band, along with a decreased participation index in the same band, are also associated with higher accuracy. In the low beta band, accuracy relates to heightened modularity and rich club connectivity during memorization. RTs are correlated with network modularity and participation index in the alpha band during memorization, as well as to average path length within the theta band and modularity in the high beta band. In the Flanker task, higher accuracy is associated with a lower clustering coefficient in the theta and alpha bands, as well as rich club connectivity in the alpha and high beta bands. When considering RT, new associations emerge with node centrality (betweenness centrality) in the theta, alpha, and low beta bands, and with nearly all measured network metrics in the low beta band.

In the resting state (Figure 5), an increase in Flanker task accuracy relates exclusively with heightened modularity within the theta frequency band. Conversely, RT is affected by an increase in average path length in the alpha and both beta frequency bands, as well as higher participation coefficients in the alpha and lower beta bands. The predictors for accuracy in the Sternberg task are more dispersed: higher accuracy is linked to a shorter average path length in the lower beta band, higher clustering coefficient in the high beta band, and higher eigenvector centrality in the alpha band. RT in the Sternberg task is associated only with a reduced clustering coefficient within the theta frequency band.

The contrast between the resting state and cognitive task (Figure 6) performance reveals a more extensive array of associations with certain cognitive measures. Specifically, accuracy in the Flanker task and RT in the Sternberg task are each associated with a limited number of network characteristics: for the Sternberg task, RT relates with clustering coefficient in the theta band and modularity in the alpha and high beta bands. In contrast, accuracy in the Sternberg task and RT in the Flanker task exhibit a broader band of associations with network reconfiguration and behavioral outcomes. Sternberg task accuracy is linked to a host of network metrics, including a shorter average path length in the low beta band, clustering coefficients across theta and both beta bands, eigenvector centrality in both the alpha and high beta bands, modularity and participation coefficients in the theta band, as well as participation and rich club coefficients in the low beta band. Meanwhile, flanker task RT, meanwhile, is associated with a reduction in average path length across all frequency bands, an increase in betweenness centrality in the theta and alpha bands, a rise in clustering coefficient in the low beta band, and a decreased participation index in the alpha band.

4. DISCUSSION

In our work, we estimated and compared the topological features of functional connectivity networks associated with performance in the Sternberg task and the Flanker task with the primary goal of identifying common and unique topological features of functional connectivity that are related to cognitive performance across two cognitive functions: WM and cognitive inhibitory control. We expected to find an overlap between the processes, reflecting a common core of cognitive functions, as well as to describe the characteristics of the unique topology specific to the two cognitive functions.

Our results indicated quite high state‐dependent specificity of network topology features. Estimation of network topology during respective tasks demonstrates that in the Sternberg task, accuracy is related to brain network dynamics within specific frequency bands, particularly in the alpha and high beta bands. The path length, modularity, and participation index within these bands emerge as significant factors, reflecting the importance of efficient global network integration and the balance between distributed and specialized processing for accurate task performance. RT in the Sternberg task, however, appears to be less dependent on specific frequency bands, indicating a wider dependence on the brain's overall network properties. Here, betweenness centrality and modularity are highlighted, indicating the need for a well‐organized network structure that can quickly adapt and reconfigure to meet the task demands. Notably, these network characteristics during the memorization phase seem to be a better predictor of both accuracy and RTs. In the Flanker task, the patterns of brain network associations with cognitive performance differ substantially from the Sternberg task. For accuracy, lower frequencies are implicated, with the clustering coefficient, modularity, and rich club coefficient being influential. This suggests that at lower frequencies, a balance between local processing and integration is essential for correct inhibition of irrelevant information. Conversely, the RT in the Flanker task is associated with higher frequency bands and depends primarily on measures of centrality. This implies that processing speed and the agility of the brain's information routing is facilitated by central nodes that likely play a pivotal role in rapid attention shifting and response execution.

These differences in the associations between network topology with accuracy and RT, observed in both tasks, may suggest that the brain mechanisms associated with these behavioral characteristics are distinct, and the characteristics themselves are not equivalent but relatively independent, each relating to different aspects of cognition.

The significant topological features estimated during the actual execution of the task differ from those derived from the resting state. Both tasks also demonstrated a substantial level of uniqueness in their respective network topologies. The Sternberg task showed unique network processes where higher clustering coefficient in the alpha band was linked to a higher accuracy and lower RTs, reduced modularity in the theta band was associated with better accuracy, and the rich club coefficient was inversely related to poorer accuracy and slower RTs across various frequency bands. For the Flanker task, unique processes included lower betweenness centrality in the alpha band, which correlates with higher accuracy, higher clustering coefficients in the beta bands linking to accuracy, and opposing associations in alpha and low beta bands for RT. Furthermore, accuracy was inversely related to eigenvector centrality in the alpha and beta bands, lower network modularity was linked to higher accuracy and longer RTs, the participation index in low beta band correlated with accuracy, and the rich club coefficient in the alpha band was inversely related to RT. Internal processes that are not directly linked to cognitive load, but rather believed to be rooted in the overall architecture of the brain, exhibit a stronger relation with the Flanker task compared to the Sternberg task. The accuracy of responses in the Flanker task is linked to network characteristics in the alpha and both beta frequency bands, whereas the accuracy of responses in the Sternberg task is predominantly associated with network measures in the theta and alpha frequency bands, which contradicts what was observed during the task itself.

While there are no overlapping network measures for accuracy between the two tasks, a commonality is observed in the processes that determine RTs: eigenvector centrality, clustering coefficient and rich club coefficient in the alpha band. However, the direction of their influence diverges; a decrease in eigenvector centrality and rich club coefficient is generally associated with longer RTs in both tasks, whereas the clustering coefficient's impact varies—a higher clustering coefficient is linked with slower responses in the Flanker task, but in the Sternberg task, a lower clustering coefficient relates to longer RTs.

For the Flanker task, the associations with RT predominantly involve a variety of network measures within the alpha band, underscoring the band's importance in task‐related processing speed. On the other hand, in the Sternberg task, the most robust association with RT is observed with the level of rich‐club organization and spans across frequency bands, indicating that the integration of high‐degree nodes into a cohesive core is vital for rapid memory processing. Comparing the networks of the resting state and during the cognitive tasks, each behavioral measure exposes a unique pattern of network activity associated with performance. The efficacy of inhibitory control, as assessed in tasks like the Flanker, may be predicated on the architecture of these preexisting networks, suggesting that the brain's baseline state sets the stage for its ability to filter and suppress distractions. Conversely, the performance in WM tasks, such as the Sternberg, seems to hinge on the brain's agility to rapidly reconfigure its networks, aligning them with the immediate requirements of the task at hand.

Some knowledge could also be gathered from the changes in frequency specificity of the resting state and cognitive task conditions. The cognitive functioning suggests an increase in higher frequency neural oscillations during task engagement, indicating that the brain is actively processing information. However, a shift is observed in the context of the Flanker task, where accuracy relates more significantly with lower frequency bands.

The reconfiguration of brain networks between the resting state and the task‐induced state is also a significant process, specific to each cognitive function. The Flanker task exhibits a stable association of higher eigenvector centrality and rich club coefficient with accuracy, and a lower average path length with RTs. This suggests that more efficient and central networks with tightly interconnected hubs contribute to better performance in tasks requiring cognitive inhibitory control. In contrast, the Sternberg task shows a primary association with the clustering coefficient, with changes in clustering between states being the sole significant predictor of RT. This indicates that for WM tasks, the global integration and the ability of brain regions to integrate with each other are critical for faster and more accurate WM performance. As both tasks exhibit mostly unique patterns of reconfiguration associated with performance, there are several shared features across the Sternberg and Flanker tasks. The impact of the clustering coefficient on accuracy is a commonality; in the Flanker task, it specifically affects the theta band, whereas for the Sternberg task, it's not restricted to a particular frequency band. Modularity in theta band also influences accuracy in both tasks, suggesting that network segregation into subnetworks is a general asset for cognitive performance. In the Sternberg task, increased clustering in the alpha and lower beta bands is associated with longer RTs. This relationship is also observed in the Flanker task, but in the opposite direction.

To summarize, our investigation has identified several neural mechanisms that persist across different cognitive scenarios. The clustering coefficient in the alpha frequency band emerges as a common factor influencing both accuracy and RTs in both the Flanker and Sternberg tasks. Modularity within the theta band and the rich club coefficient in the alpha band are linked to three out of four cognitive measures, with the exception of RT in the Flanker task. Similarly, the participation coefficient and clustering coefficient in the low beta band show relevance to three cognitive measures, excluding the RT in the Sternberg task.

Nevertheless, the majority of the connections between network measures and cognitive performance are largely distinct between the Sternberg and Flanker tasks, and are specific to either accuracy or RT. Within the Flanker task, the neural processes underpinning accuracy appear to function independently of those affecting RT. The pattern of associations suggests a transition in the frequency bands related to task performance, with accuracy associations moving from the beta band to the theta and alpha bands. Conversely, RT associations predominantly present in the alpha band during rest seem to shift toward the low beta band during the task. In contrast, the Sternberg task exhibits associations between cognitive performance measures—both accuracy and RT—and a more diverse array of frequencies and network measures, largely during the memorization phase. The Flanker task's connections are more nuanced and may be highly specific, particularly as RT for this task is nearly exclusively linked to a singular frequency band.

Our results support the notion of cognitive inhibitory control being related to the more segregated network (Cole et al., 2013; Fair et al., 2012; Rubia et al., 2014) and WM related to the more distributed network (Avery et al., 2020; Murphy et al., 2020). The Flanker task, which assesses cognitive inhibitory control, shows a different, more segregated, pattern. Task performance appears to be influenced by the individual's general brain network organization at rest, suggesting a reliance on a broader range of cognitive processes rather than a focused, task‐specific network (Cole et al., 2014). In contrast, the Sternberg task, which is heavily reliant on WM, within‐task ability to induce a broader network might be important for recruitment of stimuli‐specific brain areas (Ren et al., 2019) or particular processes in the WM (Ekman et al., 2016). RT seems to be influenced by the network's ability to quickly reconfigure and efficiently utilize its resources. Faster RTs may be facilitated by a network that can rapidly adjust its connectivity patterns (Gbadeyan et al., 2022). This dynamic reconfiguration may allow the network to bypass less efficient routes and engage in more direct, high‐speed communication channels when a rapid response is required. The efficiency of these mechanisms is likely reflected in the balance of network attributes, with certain frequency bands responsible for modulation of these processes.

Effects related to frequency bands were found for all the investigated bands. Alpha and theta bands, which are commonly associated with attention and encoding processes, are relevant to both cognitive tasks. Thus, results might be related to facilitation of efficient information retention and retrieval in WM (Herweg et al., 2020; Pavlov & Kotchoubey, 2022; Riddle et al., 2020; Wianda & Ross, 2019), top‐down control as required to resolve cognitive conflict and filter out irrelevant stimuli (Avital‐Cohen & Tsal, 2016; Buschman & Miller, 2007; Jensen & Mazaheri, 2010), and integration of information across distributed brain regions during attention‐demanding tasks (Cavanagh & Frank, 2014). Alpha oscillations are involved in attention by gating sensory inputs and suppressing non‐essential information, thus facilitating focused cognitive activity (Jensen & Mazaheri, 2010). The association of both beta band's with the Sternberg task execution suggests that higher frequency neural oscillations might be involved in the active maintenance and manipulation of information in WM (Engel & Fries, 2010; Schmidt et al., 2019).

Our data suggest that weak connections play an important role in the cognitive processes that underpin performance on the Flanker and Sternberg tasks, and this could be due to a variety of factors. The inclusion of weaker connections in the analysis has revealed additional and meaningful associations with cognitive characteristics that are typically overlooked when only stronger connections are considered. This highlights the importance of the brain's full network, including its more subtle and less prominent features, in supporting cognitive processes. In the context of the Sternberg and Flanker tasks, which measure different cognitive abilities, the weak connections seem to contribute to the nuances of cognitive performance. For example, the finding that path lengths and the reconfiguration between resting state and task states are associated with the accuracy of the Sternberg task suggests that the integration of information across different regions of the brain, facilitated by these weak connections, is crucial for successful memory encoding and retrieval (Cohen & D'Esposito, 2016; Wang et al., 2021). Similarly, the positive association of the eigenvector centrality with reaction time in the Flanker task highlights that nodes which are weakly connected to many other nodes can influence cognitive processing speed. These weakly connected nodes might serve as crucial points of filtering the incoming information, enabling noise‐free communication across the network. The negative associations observed between the clustering coefficient during memorization and weak connections across all frequency bands indicate that a more segregated network might impede the ability to store information in the brain networks. This could be due to an overreliance on local processing at the expense of broader network integration. Furthermore, the modularity's association with the memorization phase suggests that a balance between segregated and integrated processing is crucial for memory formation. High modularity, which indicates a well‐defined community structure within the network, including weaker ties, is beneficial for certain cognitive tasks.

In the Flanker task, weak connections might support the diffuse neural communication necessary for the broad integrative processes of attentional control. These weak links could facilitate the rapid reassignment of cognitive resources across the brain's network, allowing for flexible adaptation to the task's inhibitory demands (Bertolero et al., 2018). For the Sternberg task, weak connections may underlie the task's requirement for maintaining and manipulating information over short periods. These connections could be crucial for establishing transient communication channels between brain regions that are not typically strongly connected, but must coordinate during the memorization and retrieval phases (Santarnecchi et al., 2014). This might reflect the brain's strategy to engage complementary pathways to enhance memory performance, ensuring that information is not bottlenecked within a few strong connections but is diffusely supported across the network (Bassett et al., 2011). Furthermore, our findings align with the theory that weak connections contribute to the stability and flexibility of neural networks (Pajevic & Plenz, 2012). By providing additional routes for neural signaling, weak connections may endow the network with enhanced resilience against localized disruptions, ensuring sustained cognitive function even in the face of potential perturbations (X. Liu et al., 2022; Nassi & Callaway, 2009). This increased redundancy could be especially beneficial for high‐demand cognitive tasks where information transfer reliability is critical.

The network neuroscience theory (Barbey, 2018) suggests that general intelligence (g) arises from the unique topology and dynamics of the human connectome, highlighting the role of brain network architecture in enabling flexible network reconfiguration during goal‐directed behavior. It also posits that both strong and weak neural connections are crucial for intelligent information processing. Furthermore, theory distinguishes between the neural underpinnings of crystallized and fluid intelligence: crystallized intelligence relies on stable and readily accessible network states, while fluid intelligence is associated with network states that are more adaptable but less easily accessed. Our extension of this framework to additional cognitive measures generally aligns with the original concept. Both WM and cognitive inhibitory control depend on different attributes of network dynamic, indicating the brain's flexible adaptation to the unique demands of various cognitive tasks.

The Flanker task relies on more segregated networks with stable, and possibly stronger, connections, which constitute the basis for an efficient system of filtering information. Conversely, the Sternberg task is based on distributed networks that involve the engagement of weaker connections. These weaker connections may function as a dynamic and malleable network, enabling the temporary maintenance and manipulation of information. They could provide additional pathways for information flow, thus preventing overreliance on a limited set of stronger connections and ensuring a more distributed and resilient strategy for memory‐related processes (Honey et al., 2007).

A modest but constant core set of processes shared by both cognitive abilities has been identified, which most likely represents a basic network operating across several cognitive functions. This may constitute the brain's default operational mode and provide a foundation for general cognitive capabilities. This finding resonates with the notion of a common infrastructure underpinning cognitive processing (Krienen et al., 2014; Mill et al., 2017; Sadaghiani & Wirsich, 2020).

Our findings also show that WM is dependent on the same processes that underpin fluid intelligence, as described in the network neuroscience theory of intelligence (Clark et al., 2017; Shipstead et al., 2016; Unsworth et al., 2014). This suggests that WM and fluid intelligence may share common neural mechanisms, aligning with the process overlap theory (Kovacs & Conway, 2016). This theory proposes that the general factor of cognitive abilities, or “g,” relies on the overlap between cognitive processes. The presence of a stable core of processes, along with a flexible periphery that is unique to each cognitive function, might indicate that an individual's cognitive abilities depend on the reconfiguration of these peripheral processes around the central network dynamics. This conceptual framework underscores the interdependence of core and peripheral processes in determining cognitive performance, both of which are crucial for understanding the sources of individual differences in cognitive abilities.

Our study has several limitations. First, the number of intervals used to analyze brain functional connectivity was arbitrary, implying that more precise research is required to understand how cognitive performance is related to functional connectivity in the context of specific stimuli. Furthermore, examining individual trajectories of brain dynamics could be beneficial, as the dynamics during the task were not evaluated in our study. Another limitation is the lack of validation of our findings across independent datasets. While our research focused on WM and cognitive control, these areas are just a fraction of the wide spectrum of cognitive processes. Our methodology involved analyzing network construction at two threshold levels; however, the interplay between strong and weak neural connections is complex and requires further investigation. Although we applied a whole‐brain connectivity approach, cognitive performance may also be closely related to the dynamics within specific localized intrinsic brain networks. Finally, given the interrelation between WM and cognitive control, future studies should consider investigating the cross‐task relationships between functional connectivity and cognitive performance, as this could uncover intricate interactions that affect cognitive functions.

5. CONCLUSION

Overall, we investigated the neural correlates of WM and cognitive control by analyzing the Flanker and Sternberg tasks. Our findings indicate that while these tasks share some common neural mechanisms, they largely exhibit distinct network characteristics that underpin cognitive performance. The Flanker task is linked to more segregated networks, in contrast to the Sternberg task, which depends on more integrated networks, underscoring the importance of network reconfiguration in cognitive processes. Furthermore, weak connections were found to be essential for cognitive flexibility and resilience. Our results align with the network neuroscience theory of intelligence, suggesting unique network properties associated with specific cognitive functions. However, we propose that a combination of common and distinct processes is key to understanding individual differences in cognitive abilities.

CONFLICT OF INTEREST STATEMENT

The authors have no conflict of interest to declare.

Supporting information

APPENDIX S1: Supporting Information.

APPENDIX S2: Supporting Information.

APPENDIX S3: Supporting Information.

Adamovich, T. , Ismatullina, V. , Chipeeva, N. , Zakharov, I. , Feklicheva, I. , & Malykh, S. (2024). Task‐specific topology of brain networks supporting working memory and inhibition. Human Brain Mapping, 45(13), e70024. 10.1002/hbm.70024

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.