Can human brain connectivity explain verbal working memory?

Figure 1. (A) Schematic depiction of the neural network architectures and the different connectivity structures (human and monkey) used in by Schomers et al. (Citation2017). In the monkey model (left), only adjacent-areas are connected through reciprocal next-neighbour connections (black), excepted the auditory and articulatory systems, which are linked by a long distance connection (purple). In the human connectivity structure (right), next-neighbour connections are present together with “jumping links” link second-next neighbour areas (blue). (B) Connectivity matrices of the monkey and human models. The colors indicate the different kinds of links (light grey: within-area, dark grey: next-neighbor, blue: “jumping”, purple: long-distance).

Figure 2. (A) Schematic connectivity structure of the left peri-sylvian cortex in macaques, chimpanzees and humans revealed by tractography studies (modified from Rilling et al. Citation2008). (B) Structure of the neural network model with human connectivity. Superior temporal auditory (blue) and inferior-frontal articulatory (red) systems represent the peri-sylvian region. Whereas the extra-sylvian region is constituted of the lateral dorsal motor (yellow/brown) and the visual “what” (green) system. The connectivity is constituted of next-neighbour connections between adjacent areas within each system (black arrows), “jumping links” between non-adjacent areas (blue arrows) and “long distance links” between pairs of multimodal areas PB, PF_i, AT and PF_L (purple arrows, figure taken from Tomasello et al. Citation2018). (C) Schematic depiction of the neural network architecture and the different connectivity structures analysed. In simple Model (SM, left), only adjacent-areas are connected with next-neighbour connections (black) as well as long distance links between multimodal areas (purple). In the monkey Model (MM, middle) architecture, the connectivity is similar as SM and additional “jumping links” (blue) are present in the extra-sylvian region. Finally in the human (H, right) connectivity structure, “jumping links” are present in the peri-sylvian and extra-sylvian regions. (D) Connectivity matrices of the SM, MM and HM. The colours represent the different links depicted in (C).

Table 1. Main contributions of the study.

• • Regulation and control

•     • Excitatory activity was regulated using two types of inhibitory feedback, lateral local and area-specific global inhibition, to prevent non-physiologically high activity states (Braitenberg Citation1978; Knoblauch and Palm Citation2002). To prevent under-activity, each neuron received uniform, uncorrelated white noise during the learning phase. To mimic realistic variability of inputs, additional static noise was applied in the stimulus patterns supplied to primary areas (Garagnani and Pulvermüller Citation2016; Tomasello et al. Citation2018; Citation2024).

• • Cortical areas

•     • 12 areas were modelled included motor (articulatory and hand-motor) and sensory (auditory and visual) systems in the inferior and dorsolateral frontal, superior temporal and ventral temporal and occipital cortex along with multimodal areas interlinking these systems (Garagnani and Pulvermüller Citation2016; Tomasello et al. Citation2017).

• • Local connectivity

•     • Within-area connectivity was sparse, random, and initially weak along with a neighbourhood bias towards close-by links (Kaas Citation1997; Braitenberg and Schüz Citation2013).

• • Between-area global connectivity

•     • Connections between different cortical areas were bidirectional and guided by neuroanatomical principles and experimental studies, as explained in detail below.

Neuron types

Both neuron models (meanfield and spiking) follow similar membrane potential dynamics but differ in how they compute outputs. The spiking neurons capture the discrete, all-or-none firing behaviour of individual neurons, providing binary outputs based on specific activation thresholds (Connors et al. Citation1982; Matthews Citation1998). In contrast, meanfield neurons simulate the averaged behaviour of a population, producing graded responses that reflect the overall firing rate of a neural group (Wilson and Cowan Citation1973; Brunel and Hakim Citation1999). This approach allows us to explore both fine-grained and population-level neural dynamics in our simulations. The detailed mathematical equations governing each neuron type are provided in the Appendix; the model parameters were adopted from previous successful simulations (Garagnani and Pulvermüller Citation2016, Tomasello et al. Citation2017; Tomasello et al. Citation2018; Constant et al. Citation2023; Nguyen et al. Citation2024) ensuring consistency and comparability with previous established simulation work. For a more in-depth description of the parameters and their impact on CA formation, see Garagnani et al. (Citation2008).

Simulated brain areas and connectivity structure

The network modelled the following 12 cortical areas: The left-perisylvian language region was represented by six areas, including the primary auditory cortex (A1), auditory belt (PB), and modality-general parabelt areas (PB) that make up the “auditory system”. In addition, the inferior part of the primary motor cortex (M1_i), inferior premotor (PM_i), and multimodal prefrontal motor cortex (PF_i) were included, which, together, constitute the “articulatory system” (inferior face-motor and adjacent multimodal areas). Six additional areas composed the extra-sylvian region, including the primary visual cortex (V1), temporo-occipital (TO), and anterior-temporal areas (AT) for the ventral visual system. The motor system was represented by the dorsolateral fronto-central motor (M1_L), premotor (PM_L) and prefrontal cortices (PF_L). Each system included a multimodal area central to the neural architecture (PB, PF_i, AT, and PF_L). These areas can be considered “convergence zones” (Damasio Citation1989) or “connector hub areas” (van den Heuvel and Sporns Citation2013) because each of them has a relatively large number of links to other areas, and each connects 1–3 synaptic steps to all four modelled modality-specific primary cortices.

Each of the 12 areas consisted of an array of 25 × 25 (=625 in total) excitatory cells (spiking ones in the spiking model, while graded responses in the meanfield model) and 625 inhibitory cells, simulating a pool of interneurons within the same cortical columns (see , (Tomasello et al. Citation2017)); thus, each network included 15.000 neurons. Implementation of between-area connections was guided by studies using DTI/DWI and tractography (Rilling et al. Citation2008, Citation2012; Glasser and Rilling Citation2008; Thiebaut de Schotten et al. Citation2012; Rilling Citation2014; Ardesch et al. Citation2019), as described in detail below.

Taking into account the neuroanatomical differences of the AFs across primate species and human, as described in the introduction (Rilling et al. Citation2008, Citation2012; Rilling and van den Heuvel Citation2018), we constructed three distinct models: a Simple Model (SM), a Monkey Model (MM), and a Human Model (HM).

Learning phase

Prior to learning, the presence and weights of synaptic connections between neurons were determined using a stochastic algorithm with bias towards local links (Garagnani et al. Citation2007, Citation2008; Pulvermüller et al. Citation2014), using different seed values for the 12 variants of each of the 6 model types. During learning, input presentations provided co-activation of neurons. Using Hebbian learning, all weights between excitatory neurons were modified using a learning rule implementing both long-term potentiation (so that “neurons that fired together wired together”) and long-term depression (so that “neurons out of sync delinked”) (Hebb Citation1949; Artola and Singer Citation1993; Wennekers et al. Citation2006). With repeated learning, neurons that were frequently activated together strengthened their mutual connections and became more strongly interlinked than other neuron sets in the networks, which we call cell assembly circuits.

In each learning trial, a network received 16 time steps of stimulation. Stimulation patterns activated 22 randomly defined cells in 3 of the 4 primary areas of the model, whereas the 4^th primary area received random input changing from time step to time step. The 22 stimulated neurons of a pattern within each primary area of 25 × 25 cells resulted in sparse activation of around 3% of the total cells available. Stimulation patterns were applied simultaneously to primary auditory (A1) and articulatory (M1_i) areas, plus primary visual cortex (V1) to simulate learning of object-related signs or in lateral motor cortex (M1_L) for action-related signs. This mimics a word learning situation in which a novel word form is produced while a referent object is visually perceived (Vouloumanos and Werker Citation2009) or a relevant motor action is executed by the language learning infant (Tomasello and Kruger Citation1992). When learning that a word is used to speak about an object (action), hand-motor activity (visual stimulation) can be relatively variable; therefore, stimulation patterns also included uncorrelated noise pattern in V1 for action words, and in M1_L for object words, ensuring that the correlation between word form and semantic information was high in one modality but low in the non-relevant one (Tomasello et al., Citation2019; Tomasello et al., Citation2024). We note that the patterns presented to the visual and motor cortex capture experiential and environmental information relevant for learning the meaning of words – e.g., which object or action the word is used to speak about.

Each network learned 12 distinct triplets of sensory-motor patterns, simulating the learning of 6 signs thought to symbolize actions and 6 more signs related to objects. Additional white noise was always continuously presented to all areas during the learning phase, regardless of whether stimulation patterns were applied or not. After the initial 16-timestep pattern presentation, a period without pattern stimulation followed (interstimulus interval ISI). This period lasted until network activity caused by stimulation had ceased and the level of global inhibition in areas PF_i and PB had reduced to a baseline level). This was to prevent interference between consecutive patterns and learning trials. There were 5000 learning trials for each of the 12 signs (overall 60,000) for each of the 12 implementations of the 6 network types (thus 4.32 × 10^6 trials). This number of repetitions was chosen based on previous simulations and showed no substantial changes after 1000 learning episodes per sign (Garagnani et al. Citation2008; Schomers et al. Citation2017).

Evaluation

After object and action word learning, we examined the neurophysiological mechanism underpinning these processes in the three different network architectures and model types. To this end, each of the 12 learned “acoustic” patterns were presented to the auditory primary area (A1), and the resultant neural activity brought about by each of the “auditory” stimulation patterns was mapped across the entire network including all modelled areas.

Statistical analyses

To statistically test the presence of significant differences in the number of activated neurons in specific time periods and the number of CA neurons for the different model types (spiking, meanfield) and architectures (HM, MM, SM), word types (object and actions), and areas (12 cortical regions), we performed analyses using Generalized Estimating Equations (GEE). GEE is well suited for repeated measures and complex models, as it accounts for within-subject correlations and offers robust standard error estimates (Liang and Zeger Citation1986; Ballinger Citation2004; Hardin and Hilbe Citation2013). This method provides a flexible framework for the analysis of correlated data and is especially beneficial in the presence of clustered or longitudinal data, ensuring more accurate and generalizable results. The different levels of each factor were encoded with sum to 0 coding (Agresti Citation2012; Fox Citation2015), which ensured that the sum of each level was 0 and that the GEE was not biased towards one of the levels. For instance, HM was encoded as 1, MM as 0, and SM as −1. The six factors were as follows: Model Type (two levels: {Spiking, Meanfield}), Architecture (three levels: {HM, MM, SM}), WordType (two levels: {Action, Object}), and the three topological factors PeriExtra (two levels: Peri-sylvian = {A1, AB, PB, PF_i, PM_i, M1_i}, Extra-sylvian cortex = {V1, TO, AT, PF_L, PM_L, M1_L}), FrontPost (two levels: frontal ={PF_L, PM_L, M1_L, M1_i, PM_i, PF_i}, posterior={A1, AB, PB, V1, TO, AT}) and Areas (three levels: Primary = {A1, V1, M1_L, M1_i}, Secondary = {TO, AB, PM_L, PM_i} and Central = {PB, AT, PF_L, PF_i} areas). For model selection, we used the Quasi-likelihood Information Criterion (QIC) (Pan Citation2001), an extension of the AIC that is adapted to GEE models. The QIC accounts for model complexity while appropriately penalizing overfitting, making it more suitable for the type of repeated and dependent data handled by GEE. In addition to QIC, we computed the pseudo R² to evaluate the model fit and explain the proportion of variance accounted for by each model. All interactions between factors were computed and if the largest interaction was significant, the GEE was broken down in smaller models containing one factor less. In addition to be more interpretable, lower QIC and deviance were indicators that smaller models explained the data relatively better. To assess the difference in CA size between architectures, repeated measure ANOVAs were performed for each area and for each model type, followed by Bonferroni-corrected planned comparison tests (12 comparisons, corrected critical p < .0042).

To study the reverberation time (R_time), repeated one-way ANOVA (Factor: Connectivity structure (three levels: {HM, MM, SM})) was performed for each area and for each model type (spiking and meanfield). In case of significant differences in the ANOVAs, Bonferroni-corrected planned comparison tests were then carried out to identify which specific groups means differ significantly from each other. All p-value thresholds were also corrected depending on the number of comparisons using Bonferroni corrections (12 comparisons, corrected critical p < .0042).

Cell assemblies size

Across spiking and meanfield model types and across architectures, strongly interlinked cell assemblies (CAs) developed spontaneously as a result of Hebbian correlation learning (). These became manifest as the sets of neurons that jointly activated upon stimulation of an acoustic word form pattern. Intriguingly, the CAs exhibited a higher density of neural material for the fully connected human architecture (blue) and to a lesser degree for the weak connected monkey (orange) and simple (green) architectures in both spiking (right) and meanfield (left) model types. The meanfield model showed overall higher density of neural materials than the spiking model due to the different neural type models.

The distributions of the cells included in these CAs varied across network areas depending on the object and action word learning scenarios. Specifically, the extra-sylvian region showed topographical differences, in particular for the human architectures, while no differences between word types were observed in the peri-sylvian regions across architectures. The CA circuits of action words tended to extend into lateral motor regions, in particular M1_L, more strongly than those for object words, and the opposite was true for object related words in the visual system, in particular V1. These differences related to semantic circuit distributions were most obvious in the spiking human model, and they were similarly manifest in primary areas of both spiking and meanfield human model. In contrast, the other architectures, MM and SM, did not show such a clear topographical pattern distribution.

To assess the difference in CA sizes between the three architectures, Generalized Estimating Equations models and repeated-measures ANOVAs were performed on the number of cell assembly neurons in each model type obtained across all areas and for each area separately.

Simulation of auditory word comprehension: Time course activation results

After training, we simulated auditory word comprehension by presenting the learnt auditory patterns to the auditory primary area A1 and recorded the spiking neural activation of the reactivated neural circuits. This allowed us to investigate the activation time course across the different areas and models. shows the dynamics of object neural activation across the different areas for the spiking and meanfield networks and for the different connectivity architectures. Visual inspection of these activation plots reveals a main difference across model architectures. The HM showed by far the strongest neural activations for both spiking and meanfield models, outperforming both MM and SM. Object-related CA activation showed a clear activation of the visual system (V1, TO, AT) in HM for both spiking and meanfield models, with very low activation of, the dorsolateral motor system, and with some activity only the central prefrontal area (PF_L). In contrast, activation is relatively low in MM and SM. The opposite pattern was observed for action-related words with stronger activation in the motor system (M1_L, PM_L, PF_L) and less in the visual system. This is in line with the differences in CA structure described above ().

Differences in reverberation time

Here, we examined the average duration of maintained neural activity (reverberation time, R_time) after the full ignition of CAs in auditory word recognition and articulatory stimulation “trials” (). CA ignition was defined by the maximum area-specific activation. Reverberation time was defined by the number of time steps between ignition and the point in time when activity went below baseline activity. This baseline activity was computed as 2 standard deviations of activity in the 10-time step baseline before stimulation. R_time values showed clear differences across architectures but exhibited similarities between model types. Notably, in the meanfield model, the HM surpassed the MM by a factor of 2.6 and the SM by 2.9, while in the spiking model, HM reverberation time was 4.4 and 5.7 times as long as in the MM and SM, respectively. These observations were statistically confirmed by one-way ANOVAs (Factor: Connectivity structure) for both model types and for each area (corrected p-value <0.001). Below we report the results of the statistical analysis within each model types. Note, furthermore, that, in both SM and MM, reverberations were primarily present in the perisylvian network parts, whereas persistent activity in extrasylvian space was overall very weak (, middle and bottom panels). For the human model, both peri- and extrasylvian areas made a strong contribution to reverberant activity (, top panels).

CA formation across models

Word or symbol learning was implemented by concordant stimulation of the primary auditory and articulatory motor cortex of the model thought to simulate word production and perception of corresponding acoustic information, along with visual or hand-motor cortex stimulation designed to code for perceptually- or action-related features of the objects and actions to which the words relate semantically. The repeated stimulation of the same neurons led to activation spreading throughout the networks and synaptic strengthening of the connections involved, replicating well-known observations from neural networks research (Doursat and Bienenstock Citation2006; Garagnani et al. Citation2008; Lansner Citation2009; Huyck and Passmore Citation2013; Palm et al. Citation2014). As co-activated neurons were present in all network areas, the resultant CAs were distributed across network areas. Our results contrast with earlier simulations using layered networks, which arguably do not build distinct and discrete sets of neurons related to symbols, word forms, or meanings, while “represent” cognitive entities such as word forms or meanings as fully distributed dynamic activation patterns instead (Farah and McClelland Citation1991; Elman Citation1996, Citation2004; Rogers and McClelland Citation2004; Westermann et al. Citation2006; Ralph et al. Citation2017). Although distributed dynamic activation patterns can be informative to some extent, the use of more biologically well-founded models of the brain that include reciprocal topographical links between areas, excitatory within-area connections, and Hebbian correlation-based learning entails discrete circuit formation, offering a more accurate model of brain-based neural circuit formation (see (Garagnani et al. Citation2009; Pulvermüller Citation2023)).

In spite of the general observation of CA formation across the biologically constrained networks, we observed great differences in the sizes of these distributed representational units. The HMs produced CAs with substantial numbers of 50–100 cells in the spiking and 200–300 in the meanfield models. We do not interpret this quantitative difference between the two model types because no attempt was made to match activation strengths across models. The probabilistic nature of spiking activity provides an explanation why some presynaptic activations efficient in a meanfield model remained without (or very little) effect in spiking networks. This probabilistic cancelling of the effect of some activity accounts for the generally lower activations and CA sizes in the spiking models. Importantly, across network types, the human architecture led to substantially larger CAs than the monkey and simple models (see ). This difference can only be explained by the only implemented difference between these model architectures: their structural difference in connectivity. The addition of “jumping links” between non-adjacent areas, links bypassing intermediate regions, has been suggested by neuroanatomical research to be a characteristic feature distinguishing human perisylvian cortex from that of non-human primates (Rilling et al. Citation2008; Rilling and van den Heuvel Citation2018; Ardesch et al. Citation2019). These additional links imply, of course, that more activation spreads across perisylvian areas. However, over and above an activity increase, these links reduce the number of synaptic steps necessary for direct communication between distant cortical areas, thus enabling faster and more parallel neural processing. In HM, the inclusion of these jumping links during the learning phase led to enhanced synaptic strengthening within distributed cell assemblies (CAs). This strengthening was facilitated by the simultaneous activation of neurons in multiple interconnected regions, allowing for larger CA formation across different modalities. As a result, the CAs in the human-like model were not only larger in size but also exhibited higher ignition amplitudes, which means that, when activated, they generated stronger and more widespread neural responses. In contrast, models that lacked these jumping links in the perisylvian cortex (MM and SM) relied solely on next-neighbour connections there, so that information about the sounds and articulatory patterns of spoken words could only be passed sequentially from one adjacent region to the next. This led to less efficient building of neural representations for spoken words and hence to a lower CA size.

CA topographies and meaningful symbol storage

Beyond the quantitative difference in CA sizes, our results indicate qualitative differences between CAs with different functions. When simulating the learning of object-related words, the word form information was co-present with semantic information about the referent objects the words can be used to speak about. Likewise, in action word learning, information about action-related semantic features of the words was co-presented with symbol-related information to the model, mimicking early word learning as observed in language developmental studies (Tomasello and Kruger Citation1992; Vouloumanos and Werker Citation2009). The results were CA circuits that interlinked word form-related information with semantic information about referent objects and actions, respectively. Note that we simulated these differences between word types not because they capture the entire range of semantics characterizing words and symbols, but to highlight one out of many semantic differences between symbols that may affect cortical representations. Furthermore, it is needless to say that there is a wide range of different semantic word types and learning scenarios. The modelling of “pure” object and action words grounded directly into sensorimotor system is one way to address such differences. The differences in the simulated semantic grounding process, either objects or actions, led to qualitative differences of the related CAs, which extended more strongly either into the primary visual or the primary articulatory-motor cortex, respectively. However, this topographical difference in semantic grounding of object and action words was only found for the human architectures, where it was equally present in the spiking and meanfield models. In contrast, the other architectures, the monkey and simple model, did not show this semantic difference between symbol types. Only the “simple” architecture of the spiking model showed an interaction of the four factors, but this was due to an entirely different pattern of results. Word type-specific difference was confined only to extrasylvian central, connector-hub areas (AT and PF_L) and did not extend into the primary areas, as it might be expected for a grounding related difference would suggest. We do not have a ready explanation for this unexpected result in the SM but conclude that the CA topography difference related to semantic grounding in different modalities (visual vs action, i.e., V1 vs M1_L) is not directly reflected in the topographic difference of the “simple” spiking model.

In summary, our observations about the topographies of CAs representing object and action words show clearly that only the human architecture (HM) built CAs whose distribution over the network reflected the type of information relevant for the semantic grounding of symbols. A range of pre-existing neurophysiological, neuroimaging, and neuropsychological studies had previously shown category-specific semantic differences in the human brain (Warrington and Shallice Citation1984; Pulvermüller Citation2018; Grisoni et al. Citation2021; Carota et al. Citation2023), which are consistent with and may be explained by different CA topographies (Garagnani and Pulvermüller Citation2016, Tomasello et al. Citation2017; Tomasello et al. Citation2018, Citation2019). The fact that such semantic differences have model correlates, which are explained by differences in the learning paradigm, provides an argument for successful formation of meaningful symbolic representations. In this context, our present simulations suggest that the building of discrete and meaningful symbolic representations is supported by the human architecture but not by the monkey or the more “simple” ones. However, while some studies have shown that the superior temporal sulcus (STS) and superior temporal gyrus (STG) in monkeys seem to be involved in linking visual and auditory information (Joly et al. Citation2012; Froesel et al. Citation2022), they have not focused on the distinction between word modalities like action or object and do not seem to be specialized. Our results further support this point, as they showed no difference in activation patterns between these two modalities in monkeys. This suggests that central hubs in the extra-sylvian region might not be modality-specific for monkeys, unlike the human brain.

A recent study also demonstrated that macaques could reproduce a learned sequence of sounds by playing on a modified piano (Archakov et al. Citation2020). Additionally, it showed activation in the motor cortex after auditory stimulation of this sequence. This suggests that monkeys can form internal models linking auditory and motor information. In contrast, our study did not observe such a clear link between the auditory and motor systems in monkeys and simpler models for words or symbols. While higher activation provided by sequence of words could potentially reach the motor cortex in these models with further training, it is unclear if this would establish links comparable to those observed in macaques. Further research is necessary to explore this possibility.

Reverberation times and verbal working memory

Our work follows up on earlier neurocomputational work addressing the putative cortical basis of working memory (Moody et al. Citation1998; Verduzco-Flores et al. Citation2009; Pulvermüller and Garagnani Citation2014; Schomers et al. Citation2017). A clear and strong difference between model architectures was present in the duration of maintained activity within activated cell assembly circuits. The human model outperformed the monkey and simple architectures by a factor of 2.9 and 3.7 (meanfield) vs 1.8 and 2.1 (spiking model), respectively, even reaching values up to 7 when just focusing on one specific area. Generally, the longest reverberation times were present in connector hub areas central to the architectures – relative to the primary areas. The predominance of reverberation in “higher” connector hubs and association cortices sits comfortably with previous neurophysiological observations that so-called memory cells are most frequent, and show the most long-lasting sustained activity, in prefrontal and anterior temporal areas (for visual stimuli) (Fuster and Alexander Citation1971; Fuster Citation1973, Citation2022; Fuster et al. Citation1981; Goldman-Rakic Citation1999; Tomasello et al. Citation2018; Shebani et al. Citation2022), and that verbal working memory most strongly draws on perisylvian multimodal areas (Baddeley Citation2003; Fuster Citation2009; Baddeley et al. Citation2021).

Note, by the way, that the hub areas (PF_i, PF_L, PB, PM_L) did not only show the longest reverberation times in the HM, similar between-area differences were also present within both the MM and SM, although reverberation times of these models always remained substantially and statically reliably below the levels achieved by the HM. The maximal differences per area ranged between factors 7.6 (meanfield model, PM_L) and 10.1 (spiking model, PM_i) across the different architectures.

These results confirm and extend earlier findings according to which connector hub areas central to a neurocomputational architecture develop the most efficient memory mechanisms (Pulvermüller & Garagnani, Cortex Citation2014; Schomers et al. Citation2017; Tomasello et al. Citation2018; Dobler et al., Citation2024). This is explained by the relatively larger neural converging on these areas, resulting in enhanced activity levels among neurons exhibit and, consequently, the relatively larger number of neurons that are being recruited into the CA and thus take part in CA building (for discussion, see Pulvermüller & Garagnani, Cortex 2014). Note that, when comparing and , the model areas with the largest number of CA neurons also show the longest reverberation times. The biological correlates of these connector hub model areas are in the prefrontal cortex (PF areas), in the anterior temporal (area AT) and in the superior temporal gyrus and sulcus (area PB). In an extended model architecture that also includes parietal areas, posterior parietal sites would similarly be included. As these are the areas where the most pronounced memory activity has been recorded in neurophysiological experiments (Fuster et al. Citation1985, p. 198; Fuster Citation2009, Citation2022), modelling work with brain constrained networks may contribute to a mechanistic explanation of the location and cortical distribution of memory cells in primate and human cortex. This explanation, as sketched above, builds on correlation-based learning and network architecture. In this context, it is particularly interesting that not only the human architecture but also the others too exhibited the largest number of cell assembly neurons and the most prolonged memory-related activity in areas central to the architecture.

It is also important to notice that only the HM meanfield model showed differences in reverberation times between hubs areas of the extra-sylvian regions and between semantic word types, which resembled those found for the number of CA neurons. However, the meanfield model exhibited the unexpected effect of longer reverberation times for the visual-stream central areas AT for action words, and longer activation maintenance for object words in the dorsal prefrontal PF_L areas in the dorsal motor PM_L stream. The spiking model did not show this difference in central areas. Importantly, overall both models had longer reverberation times for action words in the primary motor areas and for object words in the primary visual areas.

As mentioned in the introduction, a previous neurocomputational study already found dependence of verbal working memory on human-like connectivity structure of the network. They reported that memory activity is significantly more pronounced in the human model of the perisylvian cortex, including jumping links than in a monkey architecture with next neighbour connections only. The earlier study simulated word form learning independent of the symbols’ meaning. This strategy may be criticized because, in almost all cases where infants and children learn symbol forms, they also learn what these forms mean. In the present work, we simulated the learning of word forms in the context of object perception and action execution, showing that correlation-related semantic links are built between activation patterns representing symbol forms and patterns related to referent objects and actions. This captures only quite limited aspects of referential semantics, and leaves unaddressed issues such as the great flexibility of word use across different contexts, and does not address different symbol types such as names for individual objects, terms for large conceptual categories or abstract expressions (but see (Pulvermüller et al. Citation2021; Pulvermüller Citation2023)). Still, the learning of links between symbols and referent objects and actions is an important semantic process which may be necessary for building any semantic representation. Therefore, it is well motivated to simulate referential link building in a semantic learning paradigm. The CA circuits that formed produced reverberant activity dependent on model architecture. As in the study by Schomers and colleagues, the richer human architecture outperformed the simpler models with poorer connectivity. One may argue that this difference is just due to a more or less in the number and overall strength of connections between areas and, hence, between the neurons involved. However, as shown in the earlier study, prolonging learning by a factor of 10 and therefore further strengthening of synaptic links did not significantly improve memory performance of the non-human architecture, thus indicating that connection strength may not be the only relevant factor (see in (Schomers et al. Citation2017)). A different feature distinguishing the architectures is path length, the “distance” – counted in terms of synaptic steps – between areas. In the simple and monkey architecture, it takes five steps from A1 and M1_i, whereas the shortcuts provided by the human model’s “jumping links” brings this number down to 3. Shorter pathways facilitate quicker and more direct communication between relevant areas, reducing potential delays or disruptions provided by noise in the intermediate areas (Avena-Koenigsberger et al. Citation2018). As a result, the human architecture exhibits enhanced neural communication, potentially leading to swifter and more precise processing of information that may also contribute to memory maintenance.

Our research offers a significant theoretical contribution by demonstrating how the arcuate fascicle provides the essential connections and neural shortcuts necessary for creating large cell assemblies that exhibit topographical differences important for storing semantic information for action and object words. The richer connectivity within the human language architecture enables the formation of semantic-specific neural circuits, unlike the simpler architectures in non-human models. This emphasizes the crucial role of the arcuate fascicle in supporting the semantic grounding of symbols (Harnad Citation1990), contributing to the neurobiological mechanisms that underpin verbal working memory and symbolic thought, which are unique to human cognition. By modelling these processes, the study bridges a gap in understanding how specific brain connections shape the emergence of meaningful symbol processing. From this theoretical framework, several clinical implications arise. Damage to the arcuate fascicle, as seen in conditions like conduction aphasia (Catani et al. Citation2005; Marchina et al. Citation2011) disrupts the formation and maintenance of cell assemblies crucial for producing word and sentence forms. Understanding its role in symbol formation could lead to neuroplasticity-based therapies targeting these disrupted circuits in disorders like aphasia and dyslexia, thereby improving rehabilitation outcomes (Berthier and Pulvermüller Citation2011; Berthier et al. Citation2011). The findings also suggest practical implications for artificial intelligence (AI). Many current large language models (LLMs) rely solely on text input, but our study indicates that when AI systems begin integrating multimodal inputs – such as text, vision, and motor activities – extra connections between these sensory-motor layers could enhance their ability to “understand” and process meaningful symbols (Adams et al. Citation2014). This integration could result in improved natural language understanding, bringing AI closer to human-like cognitive abilities.