Movie Features Are Hierarchically Encoded in the Dynamics of Functional Connectivity

The perception of the external environment relies on a cascade of neural computations in the visual cortex, where stimulus features are represented hierarchically. These features can be extracted through convolutional neural networks (CNNs) from naturalistic stimuli. Neuroimaging studies showed that these are processed along the dorsal and ventral visual streams and linked a gradient of increasing complexity to sequential activations and reorganization of cortical interactions. Spectrally, in the alpha and beta bands, magnetoencephalography revealed that natural viewing induced a reduction in the amplitude of functional connectivity and a maintenance of topography at rest. However, the impact of specific stimulus features on functional connectivity remains unclear. To address this gap, we acquired MEG data on 12 subjects (5 female) during movie watching and used CNN to track the dynamics of movie features. Then, we compared the temporal dynamics of CNN layers with changes in functional connectivity in the alpha and beta bands. Our findings show that feature encoding in functional connectivity relates to a complex temporal, spectral, and topological organization. In both bands, increasing stimulus complexity involved a higher number of connections and longer temporal delays between stimulus and the connectivity dynamics. Moreover, as complexity increased, we observed a shift in the involvement of connector hubs from low-level nodes to high-level ones in the visual network. Finally, a spectral dissociation emerged: the alpha band primarily encoded midlevel features, while the beta band encoded higher-level ones.