We found that long range synchronization is sensitive to working memory contents. Long-range synchronization was found to characterize VWM retention period, as had been shown in earlier work in humans using MEG (Mamashli et al., 2021; J. M. Palva et al., 2010; Sato et al., 2018; Sauseng et al., 2005). However, whether the long-range synchronization would be content-specific or related to executive functions had remained unresolved. We predicted that θ and α-band synchronization would track the contents of memory and reflect the representational functions of VWM (A. D. Baddeley et al., 2011; D’Esposito & Postle, 2015; Klingberg, 2010). Here we show that large-scale synchronization networks identified from source-reconstructed M/EEG data were modified by content and correlated with content performance. Content-specific α-band synchronization was localized into feature-selective regions within the visual systems whilst also comprising of a fronto-parietal network shared across VWM contents. Using machine learning analysis, we established that information about VWM content was greater within the long-range synchronized network than in local cortical activity. Thus, the large, distributed brain network in the α-band tracks the contents of working memory, which enables the selection of behaviorally relevant visual features to be maintained in memory, and so provides a new perspective of how working memory is organized.
Both α-band synchronization and θ-band desynchronization correlated with individual behavioral performance. High performers exhibited greater α-band synchronization and θ-band desynchronization than low performers, demonstrating the functional significance of inter-areal network synchronization in the maintenance of content-specific feature representations. The stronger α-band synchronization during the late retention period may reflect refreshing of the VWM contents, a post-consolidation process during maintenance that strengthens the memoranda and prevents loss of information (Morey & Cowan, 2018). Specifically, in the α-band, content-specific connections were localized to specific visual areas corresponding to the processing of these features (Riesenhuber & Poggio, 2000, 2002) as well as to the frontoparietal network. These findings are in line with earlier work showing that VWM contents can be tracked both in the sensory cortices and in the PFC (Christophel et al., 2017; Serences, 2016). Consistent with α-band synchronization maintaining the content of VWM, our machine learning based analysis demonstrated that these long-range connections coded for memory content. While, by contrast, the local oscillation changes provided no information about memory content. Thus, information about visual feature content i.e., the “what” in VWM is reflected in the coupling of brain areas via α-band synchronization, accompanied by decoupling via θ-band desynchronization.
The combination of α-band synchronization and θ-band desynchronization may be complementary. This could reflect the opposing demands of VWM in maintaining the internal WM representations, via α-band synchronization, and in inhibiting the external sensory stimulation (Van Ede & Nobre, 2023), achieved via θ-band desynchronization in the visual areas, which would suppress incoming new sensory information. Thus, α-band synchronization and θ-band desynchronization may operate in a complementary fashion to maintain representations and simultaneously prevent interference from incoming sensory information.
Albeit previous MEG work carried out using auditory WM (Ahveninen et al., 2023; Mamashli et al., 2021), verbal WM (Rossi et al., 2023), and LFP data in monkeys (Salazar et al., 2012) demonstrate that long-range synchronization reflects WM content, whether this synchronization could underlie human VWM has remained unresolved. Further, network synchronization could reflect a multiplexed signal capturing both the contents of WM and the executive processing underlying content selection. Using a novel graph-theory based analysis we dissociated whether network synchronization reflected shared or content-specific connections, and therefore could reflect the executive and representational sensory components of VWM, respectively. As predicted by their content specificity, both α-band synchronization and θ-band desynchronization had feature-specific connections, supporting the hypothesis that these networks reflect VWM contents. However, the networks also comprised shared connections across different memory contents, which localized to the frontoparietal control systems (Gratton et al., 2018; Power & Petersen, 2013) with the main network hubs in the IPS and FEF. This localization is in accordance with the role of α-band synchronization in top-down executive control (D’Andrea et al., 2019; Lobier et al., 2018; Mishra et al., 2021) and the role of frontoparietal regions in controlling the prioritization of information in visual WM (Sahan et al., 2020). The presence of a shared executive network across different VWM contents is analogous to the supramodal shared network across different modalities for conscious access (Sanchez et al., 2020) and suggest a similar organizational principle for WM and perception. We advance that the shared connections of α-band synchronization reflect the top-down executive network that enables the selection of the to-be-remembered visual contents, as reflected in content-specific synchronization in the visual cortices.
We found a nested relationship between local γ-band oscillations to the α-band phase as found previously (Bahramisharif et al., 2018; Griffiths et al., 2021; Lisman & Jensen, 2013; Siebenhühner et al., 2016). As the content of VWM is also reflected in local β/γ oscillations (J. M. Palva & Palva, 2018) or their bursts (Lundqvist et al., 2016), and as we show that long-range α-band oscillations reflect the top-down selection of the to-be remembered contents, we propose that the coupling of local γ-band oscillations to the α-band oscillations reflect the currently maintained spotlight on VWM contents. This neurophysiological model of VWM is in agreement with the time-based resource sharing (TBRS) model of VWM which posits an executive loop that integrates disparate representations (Barrouillet & Camos, 2015; Morey & Cowan, 2018). The concept of a network that is not a silo for a particular type of information or content would explain information leaks between memories across different types of content and allow performance to transfer or generalization between different memory types (Robertson, 2022).
We also found that working memory context could only be decoded from α-band synchronization patterns. This dovetails with the idea that α-band synchronization is responsible for maintaining the content of working memory; whilst θ-band desynchronization does not directly maintain working memory content but instead, prevents interference from ongoing sensory experiences. Earlier work has also implicated large-scale networks supporting the content of working memory (Soreq et al., 2019). In accordance with this study and with the network perspective of VWM (Christophel et al., 2017), classification accuracy for decoding VWM content was greatest when the predicted feature of each trial was decoded from α-band synchronization patterns. However, in contrast with previous work (Chen et al., 2022; Elshafei et al., 2022), the contents of working memory could not be decoded from local oscillation power.
Overall, our results establish the importance of long-range network connections for maintaining working memory contents with different frequency bands operating in a complementary fashion. WM content performance is positively correlated with α-band synchronization, while it is negatively correlated with θ-band desynchronization. This relationship may emerge because α-band synchronization maintains information about memory content while θ-band desynchronization acts to prevent interference from ongoing sensory experiences, which could be modified by executive processing. Consistent with this perspective, information about the content of working memory was revealed – using machine learning approaches – to be embedded within α-band synchronization operating across networks including visual networks that are sensitive to working memory contents. Thus, long-range synchronization may serve multiplex roles in reflecting both working memory contents and executive demands, which together provide the key architectural features of visual working memory.