Selective attention involves a feature-specific sequential release from inhibitory gating

Selective attention is a fundamental cognitive mechanism that allows our brain to preferentially process relevant sensory information, while filtering out distracting information. Attention is thought to flexibly gate the communication of irrelevant information through top-down alpha-rhythmic (8-12 Hz) functional connections, which influence early visual processing. However, the dynamic effects of top-down influence on downstream visual processing remain unknown. Here, we used electroencephalography to investigate local and network effects of selective attention while subjects attended to distinct features of identical stimuli. We found that attention-related changes in the functional brain network organization emerge shortly after stimulus onset, accompanied by an overall decrease of functional connectivity. Signatures of attentional selection were evident from a sequential release from alpha-band parietal gating in feature-selective areas. The directed connectivity paths and temporal evolution of this release from gating were consistent with the sensory effect of each feature, providing a neural basis for how visual processing quickly prioritizes relevant information in functionally specialized areas.

alpha activity is associated with the suppression of task-irrelevant visual features ( Foxe and Snyder, 2011 ;Snyder and Foxe, 2010 ; for review see Chelazzi et al., 2019 ).
Despite extensive work on the relationship between brain rhythms and inter-areal communication, the exact dynamics of attentional control and its effects at the earliest stages of visual processing remain unclear. We aimed to clarify these aspects by investigating the distributed and local effects of feature-based attention (FBA) -i.e., the selective attention mechanism that operates on the basis of low-level features of visual stimuli ( Scolari et al., 2014 ). FBA is a particularly useful paradigm to study selective attentional processing, because it involves separable regions in visual cortex that are functionally specialized for the processing of different features (feature-selective areas) and, perhaps more importantly, because FBA elicits distinct timings for the processing of taskrelevant and task-irrelevant visual information. FBA has been shown to involve the sequential and recurrent spread of activity through visual areas, in which FBA-related activations initiate rapidly in task-relevant, feature-selective areas (starting at ∼150 ms poststimulus) and propagate to other regions afterwards ( Schoenfeld et al., 2014 ). It is assumed that this sequential activation is under the control of top-down signals that flexibly regulate attention-related processing. Previously, we found that FBA is accompanied by a reactive (poststimulus) reduction in large-scale network connectivity in the alpha-band ( ∼200 ms poststimulus), which likely supports the processing of task-relevant features ( Pagnotta et al., 2020 ). Since feature-selective processing depends on attention ( Schoenfeld et al., 2014 ) and because FBA involves a network reorganization in the alpha-band ( Pagnotta et al., 2020 ), we reasoned that the previously observed cascade of activations in visual areas must be orchestrated dynamically by inhibitory top-down signals in the alphaband, from regions of the attention network. How top-down control signals evolve within tens of milliseconds after stimulus onset, and how they control the routing of activity among lower-level task-relevant areas remains an open question.
In the present study, we sought to answer this question using timeand frequency-resolved dynamic functional connectivity. We tested specific predictions derived from previous findings. First, in tasks involving selective attention, it can be expected that the posterior parietal cortex (PPC) plays a fundamental role, given its involvement in the suppression of task-irrelevant processing through top-down, alpha-band directed inhibitory signals in visual areas ( Fiebelkorn and Kastner, 2020 ;Pascucci et al., 2018 ). We, therefore, hypothesized that a fast, dynamic release from inhibitory parietal control might first occur in visual areas that are functionally specialized for the processing of attended features (task-relevant). Second, because of the highly interconnected nature of visual cortical networks ( Katzner and Weigelt, 2013 ), we hypothesized that initial changes in task-relevant areas rapidly spread to other areas through directed interactions, also to those functionally specialized for the processing of unattended features (task-irrelevant), demarcating the transition from the local to the distributed release from parietal inhibitory control across regions of the occipito-parietal cortex. The hypothesized pattern of dynamic release from parietal gating would be consistent with a neural mechanism that selectively enhances visual processing first in task-relevant areas and afterwards in task-irrelevant ones. This would explain also the previously observed sequences of enhanced neural activations, initiating in areas encoding task-relevant features, followed by areas encoding the task-irrelevant features ( Schoenfeld et al., 2014 ).
We tested these hypotheses from electroencephalography recordings (EEG) collected from healthy human subjects, in two experiments: one consisted of an FBA task on two visual features (Experiment 1), the other consisted of two sensory control tasks (Experiment 2). The results of Experiment 1 revealed a rapid, dynamic release from top-down gating influences (alpha-band) from PPC, which started with the interactions targeting task-relevant visual areas, and was followed by interactions between task-relevant and task-irrelevant areas in visual cortex. The results of Experiment 2 enabled us to determine the timing and sources of the pure sensory effect of each feature -i.e., the activation related to stimulus-driven early processing of each feature used, irrespective of attention. We found that the attention effects closely resemble the sensory effects of the two features, both from a spatial and a temporal perspective. Taken together, our results demonstrate that the release from parietal gating mediates FBA, by involving functionally specialized modules of the visual cortex that are the most sensitive to the specific features employed in this study, at latencies that are also characteristic of the features considered.

Subjects
Twenty healthy human subjects with normal or corrected-to-normal vision participated in Experiment 1 (13 females; all right-handed; ages 19-34 years, M = 23.25, SD = 4.15). Subjects' visual acuity (VA) was measured with the Freiburg visual acuity test ( Bach, 1996 ) (VA 1.10-1.63, M = 1.40, SD = 0.19). Eighteen additional subjects participated in Experiment 2 (11 females; all right-handed; ages 19-29 years, M = 23.22, SD = 2.92). In both experiments, subjects' handedness was recorded as the self-reported preferred hand for both writing and holding a computer mouse. The entire study was performed in accordance with the Declaration of Helsinki on "Medical Research Involving Human Subjects " and after approval by the responsible ethics committee (Commission cantonale d'éthique de la recherche sur l'être humain, CER-VD). All subjects participated as paid volunteers (20.-CHF/hour) and gave written informed consent prior to the experimental sessions.

Stimuli, design and procedures
In Experiment 1, subjects performed an FBA task on visual stimuli made of superimposed moving dots (Random Dot Kinematograms-RDK) and oriented Gabor patches (see below). Under different task demands, subjects either discriminated the direction of the moving dots or the off-vertical tilt angle of the Gabor. By presenting two superimposed features, differences in evoked brain activity between the two tasks (attention to Motion or Orientation) can be only ascribed to FBA, not to stimulus differences. We employed motion and orientation as features because they are known to elicit distinct activity patterns and dynamics. While orientation features are processed initially and at short latencies in the primary visual cortex (V1) ( Dupont et al., 1998 ;Koelewijn et al., 2011 ;Olshausen and Field, 1996 ), motion processing involves activity in higher-order areas (V5/hMT) at longer latencies ( Ahlfors et al., 1999 ;Schoenfeld et al., 2014 ). In Experiment 2, subjects performed sensory tasks on the same stimuli as in Experiment 1, but here we employed conditions where subjects reported the presence or absence of each given feature (i.e., the presence of a moving or static RDK in the Motionsensory task; the presence of an oriented Gabor or a noisy patch in the Orientation-sensory task; see below). In this way, we could compare i) trials with and without actual motion signal, while subjects attended to motion, and ii) trials with and without actual orientation signal, while subjects attended to orientation. This allowed separating the sensory effect of each feature from the effects of attention ( Schoenfeld et al., 2014 ).
In Experiment 1, stimuli consisted of a Gaussian-windowed sinusoidal grating (Gabor: 6°width at 3 SDs, 0.5 cycles/°spatial frequency), with superimposed Random Dot Kinematograms (RDK: 10°field-size, 1200 dots, 0.2°dot-size, infinite dot-life, 4°/s dot-speed), whose visible region was determined by the Gabor's Gaussian window ( Fig. 1 A). The off-vertical tilt of the Gabor (orientation feature) could vary between − 45°and + 45°, and its phase was varied randomly at every trial. A certain percentage of RDK dots moved coherently either towards the left or towards the right (signal-dots, motion feature), while the remaining dots (noise-dots) moved with random walks. The Gabor's tilt angle and RDK's . The same trial structure was used for both Experiment 1 and Experiment 2. The black arrow on top indicates that the RDK signal-dots are moving horizontally. In the example shown, the RDK coherent motion direction is towards right (Motion) and the Gabor off-vertical tilt is towards right (Orientation). (B-C) Stimuli used in Experiment 2 where we assessed the sensory effect of each feature in two sensory tasks: Motion-sensory (B) and Orientation-sensory (C). In the Motion-sensory task, the two trial types were 'RDK moving' (represented schematically by the black arrow on top; B-left) and 'RDK static' (represented schematically by the absence of an arrow on top; B-right), both with an oriented Gabor present. In the Orientation-sensory task, the two trial types were 'Gabor oriented' (C-left) and 'Gabor scrambled' (C-right), both in the presence of moving RDK (represented schematically by the black arrow on top). percentage of coherent motion were calibrated for each individual subject in a behavioral session to balance task difficulty across conditions (see below), and kept constant in the following neuroimaging sessions. Visual stimuli were presented on a gray background for 300 ms, around a central fixation spot that was constantly on screen (0.2°size, (0,0,160) color in RGB digital 8-bit notation), such that the stimuli were always shown at the same spatial location. Subjects were asked to perform a visual discrimination task on either the motion or the orientation feature of the same stimuli, in separate blocks. Each block of trials was preceded by an instruction indicating the relevant feature for that block (Motion or Orientation task). The inter-trial interval was randomly varied between 800 and 1000 ms, and the response interval was limited to 1500 ms (see Fig. 1 A).
All subjects participated in three separate sessions, and data were acquired in the same framework of a previously published work ( Pagnotta et al., 2020 ). The order of the three sessions was the same for all subjects. In each experimental session, task order was counterbalanced across subjects to exclude systematic effects of condition order (e.g., tiredness of the subjects). Session 1 was behavioral (approx. 45 min) and determined subject-specific thresholds for the characteristics of the two features (RDK motion coherence and Gabor orientation), to balance levels of performance (82% accuracy) across tasks in the other sessions. The behavioral session comprised two blocks (100 trials each). Subjects were instructed to report either the perceived motion direction (left vs. right) or the Gabor off-vertical tilt (left vs. right) by pressing the left or right arrows on the keyboard. Individual thresholds were derived using the adaptive staircase procedure QUEST ( Watson and Pelli, 1983 ), as implemented in PsychoPy2 ( Peirce et al., 2019 ), and psychophysical thresholds were estimated using the Weibull psychometric function (starting values: 60% for RDK motion coherence, 4°for Gabor orientation). The task-irrelevant feature was kept constant (40% RDK motion coherence in the orientation task, ± 3°Gabor orientation in the motion task). The resulting individually-calibrated thresh-olds for RDK motion coherence (5.09-86.90%, M = 37.16%, SD = 27.47%) and Gabor tilt angle (1.51-2.53°, M = 1.88°, SD = 0.29°) were used to define stimuli in the subsequent neuroimaging sessions (EEG and MRI). In the EEG session (approx. 60 min), subjects performed the two visual discrimination tasks, requiring a two-choice response (left vs. right) on either the RDK coherent motion direction or the Gabor off-vertical tilt. The EEG session comprised 160 trials per task condition, divided in 4 blocks (40 trials per block). At the end of each block, a small break was provided and its duration was autonomously controlled by each subject. In the MRI session (approx. 45 min), both functional and structural MRI data were acquired for each subject (for full details see Supplementary Materials).
In Experiment 2, two additional stimulus types were employed: one was identical to the stimulus previously used, but the dots were static (absence of motion, with Gabor still present) ( Fig. 1 B, right ); the other was obtained by scrambling the Gabor sinusoidal grating, but keeping the same low-level physical properties of the stimulus (absence of orientation, with RDK still present) ( Fig. 1 C, right ). Thresholds for RDK motion coherence and Gabor orientation were set as the averages from Experiment 1 (37.16% and 1.88°, respectively), and kept fixed for all subjects. The temporal structure of each trial remained unchanged compared to Experiment 1 ( Fig. 1 A). Subjects performed two different tasks that required a two-choice response (320 trials per task, subdivided into 8 blocks with 40 trials per block). At the beginning of each block, subjects were instructed to attend to either motion (Motion-sensory task with trial types RDK moving and RDK static; see Fig. 1 B) or to orientation (Orientation-sensory task with trial types Gabor oriented and Gabor scrambled; see Fig. 1 C), and they had to report whether the feature (moving RDK and oriented Gabor) was present or not. Task order was counterbalanced across subjects. Experiment 2 consisted of one session only, where EEG was recorded for about 70 min. In both experiments, creation and presentation of the stimuli were controlled using the Psy-choPy2 Builder ( Peirce et al., 2019 ).

EEG data acquisition and preprocessing
The acquisition of EEG data for both experiments was performed at the Department of Psychology, University of Fribourg, CH. Set-up and procedures were the same adopted in the previous study ( Pagnotta et al., 2020 ). During recording, the subject was sitting inside a dark, electromagnetic shielded room. While performing the different tasks, the subject leaned the head on a chinrest that was positioned at 71 cm from a VIEWPixx/EEG TM monitor (VPixx Technologies Inc., Saint-Bruno, CA; specifications: LCD with diagonal size of 24 inches, resolution of 1920 ×1080, luminance of 100 cd/m 2 , refresh rate of 120 Hz, and pixel response time of 1 ms). To respond during each task, the subject used a 2-button RESPONSEPixx response box (VPixx Technologies Inc.), directly connected to the compatible monitor. EEG data acquisition was performed using a 128-channel ActiveTwo EEG system (Biosemi, Amsterdam, NL), with sampling rate of 1024 Hz. In Experiment 1, subjectspecific 3D coordinates of electrodes' positions on the scalp were measured after the EEG recording, using an ELPOS system (Zebris Medical GmbH, Isny im Allgäu, DE), and these were successively used for EEG source reconstruction (see Section 2.6 ).
The same EEG preprocessing pipeline was used for both experiments. Recorded signals were downsampled to 500 Hz using an anti-aliasing filter (125 Hz cutoff frequency, 50 Hz transition bandwidth), and then the PREP plugin detrending procedure was applied using high-pass filtering with 1 Hz low-frequency cutoff ( Bigdely-Shamlo et al., 2015 ). The CleanLine adaptive filtering technique was applied to reduce line noise and higher harmonics ( https://www.nitrc.org/projects/cleanline ). The time window − 1500-1000 ms around stimulus onset was used to define epochs. Noisy EEG channels were rejected by visual inspection, and the same for epochs contaminated by noise artifacts and eye blinks occurring in the time window − 500-500 ms around stimulus onset (see below). In Experiment 1, the number of channels re-moved was in the range 10-26 across subjects ( M = 16.10, SD = 3.71) and the percentage of rejected epochs 16.67-38.75% ( M = 24.89%, SD = 5.47%); in Experiment 2, the number of removed channels was in the range 7-18 ( M = 11.44, SD = 3.55) and the percentage of rejected epochs 9.84-42.50% ( M = 23.53%, SD = 10.21%). The FastICA algorithm was then used to decompose the signals in independent components, and those identified as muscular or ocular artifacts were removed ( Hyvärinen and Oja, 2000 ). Each channel previously rejected, was interpolated using a spherical spline interpolation ( Perrin et al., 1989 ). At last, a re-referencing to the common average reference was applied to the signals ( Lehmann and Skrandies, 1980 ). EEG data preprocessing was carried out with EEGLAB and its plugins ( Delorme et al., 2011 ;Delorme and Makeig, 2004 ), in MATLAB (The MathWorks Inc., Natick, USA).
Data from one subject in Experiment 1 (male, 22 years old, 26 rejected channels and 38.75% of rejected epochs) and two subjects in Experiment 2 (2 female, ages 24 and 25 years, 15 and 18 rejected channels, 42.50% and 28.91% of rejected epochs, respectively) were excluded from successive analyses, due to excessive residual artifacts. These artifacts consisted mainly of irregular drifts over several electrodes, which were caused by inappropriate fitting of the head cap, as indicated by electrode offset outside the acceptable range during recordings (within ± 25 mV). In both experiments, trials with incorrect responses were excluded from subsequent analyses. Furthermore, for each subject in Experiment 1, the number of trials was balanced across task conditions, by randomly removing trials from the task condition that contained the most of them. This procedure was performed to avoid spurious betweenconditions differences in connectivity estimates, which may be simply due to an imbalance in the number of trials between task conditions ( Astolfi et al., 2008 ;Toppi et al., 2012 ). In Experiment 1, the maximum percentage of trials removed for each task condition was 0.63-15.63%, ( M = 6.25%, SD = 4.45%), resulting in a final number of trials per condition of 86-131 across subjects ( M = 111.16, SD = 10.40). In Experiment 2, the number of survived trials per task condition was 171-295 across subjects ( M = 246.09, SD = 34.63).
The time interval − 500-500 ms around stimulus onset was used for the successive analyses (see following sections). The poststimulus window was selected as large as possible to characterize the reactive dynamic effects of FBA, but at the same time avoiding an overlap with subjects' behavioral responses (reaction time ranged across subjects between 511.0-932.0 ms and 527.2-867.8 ms for Motion and Orientation, respectively; see Results, Section 3.1 ). The prestimulus window mirrored the poststimulus one. We also analyzed the prestimulus window to disambiguate possible effects of attentional reorienting by the subjects. Task conditions were varied block-wise to reduce these confounds and the fixation interval was randomly varied from trial to trial to avoid a fixed cognitive preparation time, but subjects are nonetheless expected to use their cognitive resources economically. Hence, they may defocus attention between the last response and before the onset of the next stimulus, even between consecutive trials of the same condition.

Statistics
In Experiment 1, all comparisons between Motion and Orientation were performed with a cluster-based permutation approach, using two-tailed dependent t -test (df = 18, p < 0.05, alpha level distributed over both tails by multiplying the p -values with a factor of 2 prior to thresholding), 50,000 permutations and p perm < 0.05 for the permutation test ( Maris and Oostenveld, 2007 ). The same settings were used in every comparison, while the space for clustering in each analysis is specified below. In Experiment 2, the cluster-based permutation approach was used to assess the sensory effects of motion and orientation (here using two-tailed dependent t -test with df = 15). Cluster-based permutation tests were performed using MATLAB codes from FieldTrip ( https://www.fieldtriptoolbox.org/tutorial/cluster_permutation_freq/ ).

EEG event-related potentials
Analyses of EEG sensor-space event-related potentials (ERPs) were restricted to three clusters of occipital electrodes, one centro-occipital (CO) and two occipito-temporal (left-lOT and right-rOT), because attention-induced modulations on early ERPs components are best detected at occipito-temporal electrode sites (e.g., see Hillyard and Anllo-Vento, 1998 ;Schoenfeld et al., 2007 ). Each electrode-cluster consisted of 7 electrode sites from the 128-channel Biosemi cap, according to the subdivision introduced by Daffner and colleagues ( Daffner et al., 2012 ). In Experiment 1, ERPs time-locked to stimulus onset were averaged separately for Motion and Orientation, and then compared between them with a cluster-based permutation approach over time frames (0-500 ms poststimulus) and electrode-clusters (CO, lOT, and rOT). In Experiment 2, ERPs time-locked to the stimulus onset were averaged separately for each of the four trial types. Like in Experiment 1, a cluster-based permutation approach was used to compare the ERPs of the two trial types of the Motion-sensory task, and the same for the Orientation-sensory task.

EEG source reconstruction
In Experiment 1, subject-specific models of i) head volume conduction, ii) sensors, and iii) cortical sources were employed to solve the forward problem. i) To create a subject-specific volume conduction model, the anatomical whole-head image of each subject was used (see Supplementary Materials). A segmentation procedure of this anatomical image was performed to obtain border surfaces between three different tissue-types (scalp, skull, and brain; smoothing with 5 voxels full-width at half maximum-FWHM Gaussian kernel), and a boundary element method (BEM) was used to create a subject-specific volume conduction model of the head ( Hamalainen and Sarvas, 1989 ). ii) To construct a subject-specific model of sensors, individual 3D electrodes' coordinates were first aligned to the corresponding head model of each subject using an interactive re-alignment procedure and then projected onto the head's surface. iii) To create a subject-specific model of sources, a template grid based on the Montreal Neurological Institute and Hospital (MNI) template anatomical MRI was employed to define the grid of solution points (1265 equivalent current dipoles, 10 mm regular spacing, constrained within cortical gray matter). Here we used the templates provided in FieldTrip ( Oostenveld et al., 2011 ) ( https://www.fieldtriptoolbox.org/template/ ). Subject-specific anatomical images were warped to the anatomical template, and the inverse operator of this procedure was applied to the template grid to derive individual grids. Each of these subject-specific grids did not possess regular spacing anymore, but in normalized MNI space it became equivalent to all the other grids across subjects, which allowed us to perform group analyses on source-reconstructed EEG data ( Pagnotta et al., 2020 ). As final step, a subject-specific lead field matrix was computed considering a forward operator with unconstrained orientation, where each solution point was modeled as three orthogonal equivalent current dipoles, placed at that location.
The linearly constrained minimum variance (LCMV) beamformer was used to solve the inverse problem ( Van Veen et al., 1997 ), where an estimate of sensor-space covariance matrix was obtained from the − 500-500 ms interval around stimulus onset. The inverse problem solution was constrained to 22 ROIs, defined using the functional data (fMRI) acquired during the MRI session, while each subject performed the visual discrimination tasks inside the scanner (for details see Supplementary Materials). A method based on singular value decomposition (SVD) was employed to compute the unique representative time series of each ROI, which best explained orientations and signal strengths of all dipoles belonging to the same ROI (for details see Rubega et al., 2019 ). Afterward, the innovations orthogonalization approach was applied to single-trial scalar-valued time series to reduce source leakage (for details see Pascual-Marqui et al., 2017 ), and consequently attenuate its detrimental effects on functional connectivity analyses ( Anzolin et al., 2019 ). The adopted methodological solutions, combined with the fact that we compared conditions at within-subject level to characterize the neural dynamic effects of feature-based attention (see Sections 2.8 -2.9 ), guarantee that the confounding effects of volume conduction on our functional connectivity results are minimal at best. EEG source reconstruction was implemented using custom MATLAB code and routines from FieldTrip ( Oostenveld et al., 2011 ) ( http://www.ru.nl/neuroimaging/fieldtrip ). For the EEG forward problem, the segmentation of MRI anatomical images was performed using the Statistical Parametric Mapping (SPM) toolbox ( Penny et al., 2011 ), version SPM12 ( https://www.fil.ion.ucl.ac.uk/spm/ ). Experiment 2 lacked individual anatomical images and electrodes' localization; therefore, EEG source reconstruction was performed using a template model of head volume conduction, sensors, and sources. To create a template head model, the same procedure as in Experiment 1 was employed (segmentation and BEM), but using directly the MNI template anatomical image. A template model of sensors was obtained from averaging coordinates of individual electrodes' positions across the subjects in Experiment 1, and this average template was aligned to the template head model. The template grid in MNI space was used as a model of sources. Like in Experiment 1, the LCMV beamformer was used to solve the inverse problem, whose solution was here obtained in all solution points rather than ROIs. The SVD-based method was employed to compute unique representative time series of each solution point, which allowed best explaining the orientation of their dipoles. To spatially localize the early sensory effect of each feature, single-trial scalar-valued time series were first averaged separately for each of the four trial types and then compared between Motion-sensory types (RDK moving vs. RDK static) and between Orientation-sensory types (Gabor oriented vs. Gabor scrambled). The comparison was performed in the earliest time interval that showed a significant sensory effect for each feature (see Results), using a cluster-based permutation approach over solution points, where neighboring points were defined based on their distance in MNI space (max distance of 15 mm for neighbors definition).

Functional connectivity estimation
Connectivity analyses were performed using the 22 ROIs sourcereconstructed, single-trial time series. The information partial directed coherence (iPDC) was employed to measure the functional directed connections between ROIs in the frequency-domain ( Takahashi et al., 2010 ). The iPDC is a measure of partialized delayed functional interactions between simultaneously recorded time-series data, based on the frequency-domain notion of Granger-Geweke causality ( Geweke, 1984 ;Granger, 1969 ;Seth et al., 2015 ). This measure has the properties of being scale-invariant and insensitive to local imbalances in signal power, and it is computationally derived from a multivariate autoregressive (MVAR) model of the time series with a preselected model order, which is the number of past observations included in the model ( Baccalá and Sameshima, 2014 ).
In this study, the Self-Tuning Optimized Kalman filter (STOK) algorithm was employed for time-varying MVAR (tvMVAR) modeling of the 22 ROIs time-series data, using a 99% variance explained criterion for setting the filtering factor threshold . To optimize within-subject model quality and to best reflect the data temporal structure, the model order for each subject was selected as the maximum optimal order across task conditions (model orders across subjects 10-22, M = 17.58, SD = 2.85), each of which was identified as the value minimizing the difference between parametric and nonparametric power spectra, respectively obtained from tvMVAR modeling and a Morlet wavelet transform (central frequency parameter 0 of 6, zeropadding to solve edge effects problem) ( Torrence and Compo, 1998 ;Pagnotta et al., 2018a ;Pagnotta et al., 2018b ). The estimation of tvM-VAR models was performed on a larger time interval around stimulus onset ( − 800-800 ms), but successive connectivity analyses fo-cused on the − 500-500 ms interval. The longer time buffers were employed for algorithmic adaptation and to accommodate that the estimates of the first time-points ( n = model order) are useless. Timevarying iPDC estimates were then computed for each subject and task condition in the 1-100 Hz frequency range (1Hz-steps). The implementation of the STOK algorithm is available on GitHub ( https://github. com/PscDavid/dynet _ toolbox ). Codes for nonparametric spectral estimation using complex Morlet wavelet are also available on GitHub ( https://github.com/mattiapagnotta/nonparametricGGC _ toolbox ).

Network analyses
The fully-connected, weighted, and directed matrix of iPDC estimates was used to define an adjacency matrix, and two types of network topology measures were estimated for each subject and task condition: global efficiency and local efficiency. The global efficiency of each network was computed as the average inverse shortest path length in the network ( Latora and Marchiori, 2001 ). The local efficiency of each node (ROI) was computed as the global efficiency of the local subgraph of that node, which is the subgraph that comprised all its closer functional neighbors, according to the principle that the stronger is the connection between two nodes, the closer their topological distance ( Latora and Marchiori, 2001 ;Rubinov and Sporns, 2010 ). At a more global level, the local efficiency of the network was then computed as the average efficiency of all its local subgraphs ( Rubinov and Sporns, 2010 ). Graph measures were derived using the Brain Connectivity Toolbox ( Rubinov and Sporns, 2010 ) ( http://www.brain-connectivity-toolbox.net ).
Selective attention is mediated by neural mechanisms that promote inter-areal communication through specific rhythms, to convey relevant signals in cortical large-scale networks ( Bonnefond et al., 2017 ;Buschman and Kastner, 2015 ;Spitzer and Haegens, 2017 ). These functional networks should therefore be characterized by a topology that facilitates communication and information routing among cortical areas. This can be achieved through a network organization that favors both functional segregation (i.e., specialized processing) and integration (i.e., access/combine specialized information) ( Rubinov and Sporns, 2010 ). We used the local and global efficiency as topological measures of functional segregation and integration, respectively ( Latora and Marchiori, 2001 ;Rubinov and Sporns, 2010 ), providing a time-and frequency-resolved analysis of how information routing among regions of interest (ROIs) varies under FBA. In iPDC-based graphs, local and global efficiency can respectively quantify the levels of functional segregation and global functional integration among ROIs (for details see Pagnotta et al., 2020 ).
To assess the local changes in network topology due to FBA, the timevarying and frequency-specific estimates of single-node local efficiency were compared between Motion and Orientation, using a cluster-based permutation approach over time frames ( − 500-500 ms around stimulus onset), frequencies (1-100 Hz) and all ROIs. Effect sizes of significant differences were estimated using Cohen's d ( Cohen, 1992 ), after averaging estimates in the time-window 16-88 ms and frequency range 8-24 Hz, separately for each task condition. These intervals of interest for time and frequency were selected in a data-driven way as FWHM of the frequency-collapsed time distribution and of the time-collapsed frequency distribution of significant between-conditions differences, respectively (see Results and Fig. 3 A). An analogous approach was employed to separately compare the network's global and local efficiencies, over time frames ( − 500-500 ms) and frequencies (1-100 Hz).
Since we found between-conditions differences in single-node local efficiency (see Fig. 3 ), a measure of Relative Graph-measure Change (RGC) w.r.t. baseline was estimated, separately for each condition and in each ROI where we observed local efficiency differences, to assess whether these differences were due to a reactive increase in Motion or a decrease in Orientation (or a combination of the two). RGC was estimated by subtracting and normalizing the time-frequency local efficiency estimates by the frequency-resolved baseline average ( − 300-0 ms prestimulus). To assess whether RGC differed significantly from baseline in the frequency range of interest 8-24 Hz (see above), the following bootstrap approach was employed: a distribution of across subjects RGCs (after averaging in the frequency range) was obtained by resampling with replacement ( n = 10,000); 95% confidence intervals (CIs) from this bootstrap distribution were estimated using the biascorrected and accelerated (BCa) method ( Efron, 1987 ;Efron and Tibshirani, 1993 ); variations in poststimulus dynamics were identified at time instants when the CI lower bound was above zero (significant increase) or the CI upper bound was below zero (significant decrease).

Dynamic connectivity differences
Connectivity differences between Motion and Orientation were assessed on individual connections, focusing on the intervals of interest for time and frequency (16-88 ms, 8-24 Hz), based on the network topology results. The frequency interval was subdivided into the two canonical bands for analysis (alpha: 8-12 Hz; beta: 13-24 Hz). In each frequency band, time-varying connectivity estimates were collapsed over frequencies and time points in the window, by averaging iPDC estimates separately for each task condition. Differences between Motion and Orientation were then assessed using a cluster-based permutation approach over all connections between ROIs. In each band, a measure of Relative Connectivity Change (RCC) w.r.t. baseline, was obtained by subtraction and normalization of the time-varying connectivity estimates by baseline average ( − 300-0 ms), separately for each condition and in each connection where we found between-conditions differences (see Results and Fig. 5 ). To assess whether within-condition alpha-band connectivity decreases followed a specific pattern, their timing was assessed in the 0-300 ms poststimulus window, by estimating the first time interval of decreased connectivity on all connections between parietal, visual, and temporal ROIs in the right hemisphere, using a bootstrap approach. For each connection, a bootstrap distribution of across subjects RCCs was obtained by resampling with replacement ( n = 10,000), and the BCa method was used to compute 95% CIs from this bootstrap distribution ( Efron, 1987 ;Efron and Tibshirani, 1993 ). A significant decrease in connectivity was identified when the upper bound of the BCa 95% CI was inferior to zero.

Event-related potentials
The aim of Experiment 1 was to assess the neural dynamics of FBAinduced modulations. After calibrating task difficulty in a preliminary experimental session (see Methods, Section 2.2 ), subjects performed the two tasks with comparable accuracy (percentage correct in the Using nonparametric permutation testing ( Maris and Oostenveld, 2007 ), based on two-tailed dependent t -test ( p < 0.05) and with p perm < 0.05 for the permutation test, we compared ERPs, time-locked to stimulus onset, between Motion and Orientation (see Methods, Section 2.4 ). The comparison was performed in three groups of occipital electrodes where attentional modulations are typically observed ( Daffner et al., 2012 ): left occipito-temporal (lOT), centro-occipital (CO), and right occipito-temporal (rOT) ( Fig. 2 ). In CO, we found a significant ERP increase for Orientation compared to Motion (74-90 ms; p perm = 0.0476; Fig. 2 B), due to a larger P1 component (positive ∼100 ms) for Orientation, as revealed by looking at the ERP polarity in each condition. In rOT, we found significant differences (Motion higher than Orien-

Fig. 2.
Sensor-space event-related potentials. Grand-average ERPs for Motion (red) and Orientation (blue) are shown in three electrode clusters: left occipitotemporal (A), centro-occipital (B), and right occipito-temporal (C). Each electrode cluster is highlighted in gray in the topography plot on the left, which represents the 128-electrode layout. For each condition, the shading represents the standard error of the mean. In each plot, gray vertical shades represent latencies that showed statistically significant differences between task conditions. tation) at 154-176 ms ( p perm = 0.0291) and 194-274 ms ( p perm = 0.0015; Fig. 2 C). By looking at the polarities of ERP components in each condition, we observed that the differences at 154-176 ms were due to a larger N1 component (negative ∼170 ms) for Orientation than Motion, vice versa those at 194-274 ms were due to a larger P2 component (positive ∼200 ms) for Motion ( Fig. 2 C). The analysis of three electrode-groups allowed us to show that the early FBA effects were lateralized, involving predominantly right and central occipital electrode sites. In lOT, on the other hand, we observed significant ERP differences only at longer latencies, after 450 ms (Motion higher than Orientation; p perm < 0.001; Fig. 2 A). The results showed analogous late modulations (after 450 ms) in rOT ( p perm < 0.001; Fig. 2 C) and CO ( p perm = 0.0347; Fig. 2 B), where we also found increased ERP for Orientation ∼400 ms poststimulus ( p perm = 0.0024; Fig. 2 B).
While the differences observed after ∼400 ms poststimulus likely reflect late mechanisms of processing (e.g., decisional), the early FBA effects show that attentional selection affects perceptual processing shortly after stimulus onset. Together our results reveal larger ERP components at shorter latencies when orientation is attended ( ∼P1-N1) and at longer latencies when motion is attended ( ∼P2), indicating that attention to orientation selectively enhances neural activations earlier than what attention to motion does. These distinct timings are consistent with previous observations for the sensory processing of orientation and motion features ( Ahlfors et al., 1999 ;Dupont et al., 1998 ;Koelewijn et al., 2011 ;Olshausen and Field, 1996 ;Schoenfeld et al., 2014 ). Our results reveal also a right lateralization of these early FBA effects in sensor-space; however, these do not directly allow any meaningful interpretation in terms of underlying cortical sources. For this reason, we employed source reconstruction techniques to estimate source-reconstructed signals in cortical ROIs (see Methods, Section 2.6 ), and the successive connectivity analyses were performed on sourcespace (see below).

Feature-specific topological network differences
We employed two graph theory measures, local and global efficiency, to quantify the large-scale functional network organization and assess network changes induced by feature selection (see Methods, Section 2.8 ). Nonparametric permutation testing ( Maris and Oostenveld, 2007 ) was used to assess FBA-induced differences in network topology. We found significant attentional differences in local efficiency in the right inferior parietal lobule (IPL; p perm = 0.0464), primary visual cortex (V1; p perm = 0.0297), and secondary visual cortex (V2; p perm = 0.0245). These differences were distributed over latencies between − 28 and 124 ms around stimulus onset, at frequencies from 1 to 36 Hz, and showed increased local efficiency for Motion compared to Orientation ( Fig. 3 A), with a medium effect size (Cohen's d ranging between d = 0.38 (in V2) and d = 0.46 (in IPL) across ROIs; Fig. 3 B) ( Cohen, 1992 ). Additionally, we found a trend for higher local efficiency for Motion than Orientation in right visual area V5 ( p perm = 0.0850; d = 0.32), which is known to encode motion coherence ( Ahlfors et al., 1999 ;Schoenfeld et al., 2014 ). We did not find reliable effects of FBA on the global efficiency of the network (see Supplementary Materials, Fig. S2).
The observed between-conditions differences in ROIs' local efficiency were due to a decrease in efficiency for Orientation, rather than an increase for Motion, as evidenced by examining the time-varying and frequency-specific RGC (see Supplementary Materials, Fig. S3). This was confirmed using a bootstrap approach, which allowed to assess withincondition dynamic differences in local efficiency compared to baseline, in the frequency range 8-24 Hz (see Methods, Section 2.8 ). We found significant dynamic decreases for Motion in all four ROIs ( Fig. 4 A): in IPL sustained after 336 ms, in V1 at 92-104 ms and sustained after 152 ms, in V2 at 180-191 ms and sustained after 236 ms, and in V5 sustained after 168 ms. For Orientation, significant dynamic decreases were generally observed at shorter latencies ( Fig. 4 B): in IPL at 28-92 ms and nearly sustained after 304 ms, in V1 at 0-68 ms and sustained after 184 ms, in V2 at 28-100 ms and sustained after 172 ms, and in V5 at 40-100 ms and nearly sustained after 192 ms. We did not find any significant dynamic increase in ROIs' local efficiency. These results show two network topology effects of attentional selection: attention to orientation decreases the local efficiency of right parietal and occipital ROIs in the alpha/beta-band, at ∼0-100 ms, likely reducing communication from these areas; while attention to motion leads to a similar decrease of local efficiency, in the same ROIs but at longer latencies (after ∼150 ms poststimulus, except in V1 where a first significant decrease was observed ∼100 ms). Taken together, the network topology results show that FBA involves changes in functional communication and information routing in the alpha and beta-band, which are orchestrated by right parietal and occipital cortex. These findings further reveal that the neural mechanisms underlying the early FBA-induced ERP differences (see Fig. 2 ) are a product of feature-specific changes in network topology, rather than simply reflecting changes in the local neural activity.

Individual connections contributing to topology differences
Having identified the latencies and frequencies at which FBA induced changes in network topology (see Fig. 3 ), we next asked which individual functional connections contributed to this topological reorganization, separately for the alpha (8-12 Hz) and beta-band (13-24 Hz) (see Methods, Section 2.9 ). Using nonparametric statistical testing ( Maris and Oostenveld, 2007 ), in the alpha-band, we found higher connectivity strengths for Motion than Orientation, for the directed connections from IPL to superior parietal lobule (SPL; p perm = 0.0248; d = 0.53), SPL to V1 ( p perm = 0.0174; d = 0.78), and V1 to middle temporal gyrus (MTG; p perm = 0.0400; d = 0.54), all in the right hemisphere ( Fig. 5 A). These attentional differences were mainly driven by a decrease in connectivity strength for Orientation relative to baseline, in the first 100 ms poststimulus, while such decrease was generally less pronounced for Motion (see Supplementary Materials, Fig. S4A). We did not find higher connectivity strengths during Orientation compared to Motion. It is worth highlighting that the observed alpha-band modulations are right-lateralized, therefore they could underlie the early FBA effects observed previously in the sensor-space analysis (see Fig. 2 ). Together these results reveal the presence of FBA-induced changes on the path of directed connections from IPL and SPL (PPC) to V1 (topdown influences), and from V1 to MTG (visual cortex connection). Considering i) the connections involved, ii) the observed tendency for an early poststimulus decrease in connectivity when orientation was attended, and iii) the fact that alpha is the candidate neural rhythm for parietal gating signals, FBA control may be exerted through a selective release from alpha-driven gating in visual cortex. The timing of such release may also be feature-specific, as indicated by our sensor-space ERP results (see Fig. 2 ). To test this, we next assessed the latencies of within-condition poststimulus decreases in alpha-band connectivity, statistically compared to baseline (see Section 3.4 ).
In the beta-band, we also found increased connectivity for Motion compared to Orientation, for the connections from right Fig. 4. Within-condition Relative Graph-measure Change (RGC) in the alpha/beta-band. Grand-average RGC for the local efficiency is shown in each ROI that showed between-conditions differences in the alpha/beta-band (8-24 Hz), separately for each task condition: Motion (A, dark red) and Orientation (B, dark blue). In each plot, the shading represents 95% CIs and the gray vertical shades represent latencies that showed statistically significant differences compared to baseline (increase or decrease) from the bootstrap approach. MNI template with superimposed significant beta-band connectivity differences. In each plot, significant between-conditions differences are represented by arrows (red indicates Motion higher than Orientation; blue indicates Orientation higher than Motion). The absence of blue arrows indicates that there were no significant effects in that direction.
premotor area (PM) to right V1 ( p perm = 0.0250; d = 0.56), left PM to left SPL ( p perm = 0.0086; d = 0.50), right superior temporal gyrus (STG) to left SPL ( p perm = 0.0084; d = 0.32), and for the interhemispheric connection from right to left IPL ( p perm = 0.0175; d = 0.66) ( Fig. 5 B). These attentional differences were driven by decreased connectivity for Orientation in the case of the left PM to left SPL connection, while dynamic changes were absent for the right PM to V1 connection. For the remaining connections, differences were due to increased connectivity strength for Motion (see Supplementary Materials, Fig. S4B). In the beta-band, we did not find any significant increase for Orientation compared to Motion.

Feature-specific patterns of release from gating
Our results indicated that dynamic decreases in alpha-band connectivity may play a key role in mediating FBA, quickly spreading over multiple connections. To identify whether a characteristic pattern of release from alpha-band inhibition exists for attention to orientation or attention to motion, we evaluated the timing of statistically significant poststimulus decreases in alpha-band connectivity strengths, on all connections between parietal, visual, and temporal ROIs in the right hemisphere (see Methods, Section 2.9 ). We found that attention to orientation decreased alpha-band connectivity first from IPL to SPL, from V1 to V5/MTG ( ∼20-50 ms poststimulus), and then from V5 to V1 (feedbacklike) and from V2 to STG ( ∼60 ms poststimulus; Fig. 6 A-B). Attention to motion, instead, decreased alpha-band connectivity first from IPL to MTG ( ∼125 ms poststimulus), then from SPL to V5, to V1, and back to V5, and together from V1 to V2 ( Fig. 6 C-D). In both task conditions, we then found a general spreading of release from alpha-band influences on other connections, especially among visual and temporal areas ( Fig. 6 A-C). Together, these results reveal feature-specific, dynamic patterns of release from alpha-band gating influences, each of which starts with the connections from PPC to task-relevant areas and between task-relevant and task-irrelevant areas.

Sensory effects of motion and orientation
The dynamics of the release from alpha-band influences closely resembled the sequential activation expected during sensory processing (i.e., activations in V1 for orientation and in V5/hMT for motion processing) ( Ahlfors et al., 1999 ;Dupont et al., 1998 ;Koelewijn et al., 2011 ;Olshausen and Field, 1996 ;Schoenfeld et al., 2014 ). To confirm this correspondence under the paradigm used here, we performed a second experiment. Experiment 2 enabled us to isolate the sensory effect of each feature (see Methods, Section 2.2 ), which allowed verifying whether the location of regions that were firstly released from top-down alphaband inhibition (see Fig. 6 ) corresponded to the location of regions initially devoted to the processing of such features. Using nonparametric statistical testing ( Maris and Oostenveld, 2007 ), we compared ERPs time-locked to stimulus onset to identify the temporal sensory effect of each stimulus feature (see Methods, Section 2.5 ). For the Motion-sensory task, significant scalp ERP differences were observed in lOT (158-368 ms; p perm < 0.0001), CO (106-142 ms and 158-382 ms; p perm = 0.0131 and p perm = 0.0002, respectively), and rOT (162-310 ms; p perm = 0.0013) ( Fig. 7 A). For the Orientation-sensory task, significant scalp ERP differences were found in lOT (150-194 ms and 230-270 ms; p perm = 0.0059 and p perm = 0.0180, respectively), CO (68-90 ms, 94-130 ms, and 146-206 ms; p perm = 0.0392, p perm = 0.0099, and p perm = 0.0004, respectively), and rOT (84-104 ms and 150-266 ms; p perm = 0.0478 and p perm = 0.0002, respectively) ( Fig. 7 B). These results demonstrate that, with these stim-uli and tasks, the sensory orientation effect starts as early as ∼70 ms poststimulus, while the earliest sensory effect of motion is ∼110 ms poststimulus.
To identify the sources of these sensory effects, we performed EEG source reconstruction and compared ERPs in source-space, focusing on the first time interval showing a statistically significant sensory effect for each feature (see Methods, Section 2.6 ). The sources of the sensory effect for motion were localized in the right middle temporal cortex ( Fig. 8 A), while those for orientation were localized in the central occipital and parietal cortex ( Fig. 8 B). These sources were located in the proximity of the target visual ROIs that showed the initial decrease of alpha-band directed connectivity from IPL and SPL, in the right hemisphere (see Fig. 6 ): orientation-specific area V1 and motion-specific areas V5 and MTG. The results demonstrate that the feature-specific patterns of release from alpha-band inhibition, observed in Experiment 1, are consistent with the timing and cortical sources of the sensory effect of each feature.

Discussion
Selective attention constantly influences perception, prioritizing the processing of what is relevant over what is not. The brain mechanisms underlying attention, and particularly FBA, are distributed and involve the selective activation of feature-selective regions in primary sensory cortices, as well as the top-down, regulatory control from higher-level Fig. 7. Sensory effect of each feature: sensor-space ERPs. (A) Grand-average ERPs for the Motion-sensory task (trial types: RDK moving in red and RDK static in orange) are shown in three electrode clusters: left occipito-temporal (top), centro-occipital (middle), and right occipito-temporal (bottom). (B) Grand-average ERPs for the Orientation-sensory task (trial types: Gabor oriented in blue and Gabor scrambled in light-blue) are shown in three electrode clusters: left occipito-temporal (top), centro-occipital (middle), and right occipito-temporal (bottom). Electrode clusters are highlighted in gray in the corresponding topography plots on the left, which schematically represent the 128-electrode Biosemi scheme. For each trial type, the shading represents the standard error of the mean. In each plot, gray vertical shades represent latencies that showed statistically significant differences between task conditions. regions. Our findings revealed that FBA unfolds rapidly through the sequential reorganization of directed functional interactions mediated by the right PPC. We found that the right PPC acts as a gatekeeper for attentional processing, by selectively enabling or down-regulating activity in downstream areas through influences in the alpha-band. These influences changed quickly during feature selection, with paths and dynamics that depended on the functional specialization of regions along the visual stream. To the best of our knowledge, these findings are the first to show a dynamic connectivity mechanism of prioritization during feature selection, which is flexibly controlled by the right PPC through top-down alpha-band connections to feature-selective visual areas. Our findings show that stereotyped patterns of release from alpha-driven gating mediate the selection of behaviorally relevant features during FBA.

Top-down control through alpha influences
We found that FBA involves a network reorganization in the alpha and beta-band, which locally affects areas in the right parietal and occipital cortices (see Fig. 3 ). We revealed that these FBA effects were driven by poststimulus decreases in local efficiency (see Fig. 4 ), characterized by distinct timings for attention to orientation ( ∼0-100 ms and after ∼172 ms) and attention to motion (after ∼150 ms poststimulus). By analyzing FBA-induced changes on individual connections, we observed that the alpha-band topological reorganization was due to a decrease in the influences from IPL and SPL to V1, and on the connection from V1 to MTG, when attending orientation (see Fig. 5 A). For each task condition separately, we then assessed the poststimulus dynamic decreases in alpha-band connectivity and showed that a characteristic pattern of release from alpha-band parietal connections exists when attention is directed to each feature (see Fig. 6 ). Previous studies have shown that alpha-band signals from the parietal cortex gate local processing in visual areas, preventing the propagation of behaviorally irrelevant activity ( Fiebelkorn and Kastner, 2020 ;Pascucci et al., 2018 ). Conversely, local processing and information routing are facilitated when visual areas are released from the inhibitory parietal gating ( Bonnefond et al., 2017 ;Jensen and Mazaheri, 2010 ). Our findings are the first to show that the release from alpha-driven gating is dynamic, occurring first in feature-selective areas, at feature-specific latencies, and then spreading through a network of occipital and temporal regions in an orderly way.
The observed release from parietal gating is consistent with the functional specialization of visual areas and with the selective enhancement of local processing in task-relevant areas. When orientation was attended, the release from gating involved first the directed connections from parietal areas (IPL, SPL) to orientation-specific areas (V1), and from V1 to motion-specific areas (V5, MTG). These modulations are consistent with the enhancement of local processing in V1, which is functionally specialized for orientation features ( Dupont et al., 1998 ;Koelewijn et al., 2011 ;Olshausen and Field, 1996 ), and the presence of a release from alpha-band inhibition that spreads from V1 to areas specialized for processing motion (i.e., the task-irrelevant feature) ( Ahlfors et al., 1999 ;Schoenfeld et al., 2014 ). When motion was attended, the pattern was reversed, involving the connection from IPL to MTG and the path from SPL to V5 to V1. Our results provide evidence that a specific neural mechanism, the release from inhibition ( Bonnefond et al., 2017 ;Foxe et al., 1998 ;Jensen and Mazaheri, 2010 ;Mazaheri and Jensen, 2010 ), supports FBA along the hierarchy of functionally specialized modules of the visual cortex ( Ahlfors et al., 1999 ;Dupont et al., 1998 ;Koelewijn et al., 2011 ;Olshausen and Field, 1996 ;Schoenfeld et al., 2014 ). This conclusion is corroborated by the results of our second experiment, where we found a close correspondence between the initial targets of the release from parietal gating and the first regions involved in the early sensory processing of the visual features (see Fig. 8 ).

The involvement of other neural rhythms
A proposed mechanism for how top-down signals mediate local processing in target areas is cross-frequency coupling, with lower frequencies co-determining the high-frequency activity within areas ( Buzsáki, 2006 ;Canolty and Knight, 2010 ;Jensen and Colgin, 2007 ). The nested oscillations framework proposes that the synchronization of alpha rhythm supports inter-areal communication, by aligning faster rhythms (gamma, above ∼30 Hz) associated with local-circuit computations with the peaks of alpha driving. This way, alpha-gamma coupling can coordinate the information routing of local activity among distinct regions in time ( Bonnefond et al., 2017 ). In visual attention tasks, increased alpha-gamma coupling has been observed over visual cortices ( Voytek et al., 2010 ) and in feature-selective areas during FBA ( Pagnotta et al., 2020 ). These coupling increases are typically accompanied by decreased alpha power in occipital areas and decreased network efficiency ( Pagnotta et al., 2020 ;Pascucci et al., 2018 ). One of these studies has furthermore shown an inverse relationship between alpha-band top-down signals from attentional control areas and alphagamma coupling in occipital areas, demonstrating that alpha-band inhibitory signals from the parietal cortex disrupt local processing and information routing at target regions ( Pascucci et al., 2018 ). In this view, attention-induced releases from inhibition could promote communication and facilitate the propagation of behaviorally relevant information ( Bonnefond et al., 2017 ;Jensen and Mazaheri, 2010 ). Our current findings provide empirical support for this perspective and demonstrate that FBA operates through a stereotyped pattern of dynamic release from inhibitory gating.
We observed a large-scale network reorganization induced by FBA, determined by significant poststimulus decreases in local efficiency for each attended feature (see Fig. 4 ). The local efficiency of each ROI is a measure of specialized processing in that ROI, which quantifies the level of functional integration within the ROI-specific subgraph of interconnected regions ( Latora and Marchiori, 2001 ;Rubinov and Sporns, 2010 ). Hence, our findings suggest that FBA reduces communication among right parietal and occipital areas at low frequencies; furthermore they extend previous attention studies ( Pagnotta et al., 2020 ;Pascucci et al., 2018 ), by showing that this reduction happens at different latencies depending on the attended feature (earlier for orientation, later for motion). Although we did not find FBA differences at higher frequencies (gamma), we observed a tendency for poststimulus increases in gammaband efficiency, for both task conditions (see Supplementary Materials, Fig. S3). This is in line with our previous report of an inverse alpha-gamma relationship for large-scale functional network organization ( Pagnotta et al., 2020 ), and with previous studies documenting an inverse relationship between alpha-band signals and gamma-band activity ( Bonnefond and Jensen, 2015 ;Mathewson et al., 2011 ;Mazaheri and Jensen, 2010 ;Pascucci et al., 2018 ).
The FBA-induced network reorganization involved also frequencies in the beta-band. The connections that contributed the most to this reorganization were between prefrontal and parietal areas (see Fig. 5 B). Contrary to top-down alpha-band signals that mediate attentional control in the presence of external visual stimulation, modulations in beta-band influences have been associated with endogenous mechanisms that operate predominantly before stimulus onset, in anticipation of task-relevant information (for review see Spitzer and Haegens, 2017 ). Our results suggest that when attending motion such influences are maintained after stimulus onset (see Supplementary Materials, Fig. S4B), reflecting the persistence of a motion-specific preparatory stage of processing. For attention to orientation, we observed instead a disengagement of the betaband influence from prefrontal to parietal cortex around stimulus onset, possibly indicating a switch in processing from endogenously controlled to stimulus-driven.

The role of PPC as gatekeeper
We hypothesized that PPC acts as a gatekeeper for attentional processing, through top-down directed signals in visual areas. Taken together, our connectivity results confirmed this hypothesis, revealing the presence of an attentional modulation by which the right PPC prioritizes the local processing in task-relevant over task-irrelevant visual areas, through a selective release from alpha-driven gating. This supports and extends the findings of several previous fMRI studies that showed that PPC is a domain-independent, common hub for attentional control ( Wojciulik and Kanwisher, 1999 ; for reviews see Corbetta and Shulman, 2002 ;Corbetta et al., 2008 ). PPC activity generalizes in fact across a wide variety of attention tasks: when subjects attend visual stimuli based on their spatial locations and during voluntary shifts of attention between locations ( Corbetta et al., 2000 ;Hopfinger et al., 2000 ;Yantis et al., 2002 ), when subjects perform object-based selections ( Serences et al., 2004 ), and in studies of FBA ( Greenberg et al., 2010 ;Liu et al., 2003 ). Our results also confirmed the frequency-specificity of the top-down influences from PPC to visual areas ( Fiebelkorn and Kastner, 2020 ;Pascucci et al., 2018 ) (see also Section 4.1 ). Notably, anatomical evidence that these connections from PPC to visual cortex mediate attentional signals has been previously provided using whitematter imaging and fMRI ( Greenberg et al., 2012 ).

Feature-specific timings for stereotyped patterns of release from gating
We showed that FBA involves stereotyped patterns of dynamic, poststimulus release from alpha-band influences (see Fig. 6 ). These feature-specific patterns are consistent with the functional specialization of regions in visual cortex (see Section 4.1 ), as well as with the expected timing of processing of each feature ( Ahlfors et al., 1999 ;Dupont et al., 1998 ;Koelewijn et al., 2011 ;Olshausen and Field, 1996 ;Schoenfeld et al., 2014 ). The initial targets and latencies of the release depended on the attended feature. For orientation, the effects occurred at earlier latencies ( ∼20-50 ms poststimulus) than for motion features ( ∼125-180 ms). The selection of visual features has previously been shown to involve a temporal sequence, in which neural activity is enhanced first in cortical areas encoding the relevant feature and then in areas encoding the irrelevant feature ( Schoenfeld et al., 2014 ). Our findings extend this view by revealing the dynamics of inter-areal functional interactions during feature selection. Specifically, we demonstrated a reversal in the pattern of release from alpha-driven gating between attention to orientation and attention to motion. It is worth mentioning that Schoenfeld and colleagues used color and motion as visual features, localizing the color-specific area in the inferior occipital cortex and the motion-specific area in V5/hMT, and for both features, they showed sensory effects starting ∼120 ms poststimulus (lasting over the time range ∼120-250 ms) ( Schoenfeld et al., 2014 ). Here we used orientation and motion features instead, and showed that their sensory effects start respectively ∼70 ms and ∼110 ms poststimulus (Experiment 2). This allowed us to separate the selection of each feature not only by cortical area (orientation processing in V1, motion processing in V5/MTG) but also by their timing (orientation processing earlier than motion processing).
For attention to motion, we found enhanced neural activations compared to attention to orientation ∼200 ms poststimulus, as revealed by significant between-conditions ERP differences (see Fig. 2 C). Furthermore, using the local efficiency of each ROI, we found significant poststimulus decreases after ∼150 ms in right parietal and occipital areas (IPL, V2, and V5), and ∼100 ms and after ∼150 ms poststimulus in V1 (see Fig. 4 A). We then found significantly decreased alpha-band connectivity strengths at ∼125-180 ms poststimulus (see Fig. 6 C). Hence, for the motion-specific dynamics, the timing of attention effects matched the timing of the pure sensory effect (irrespective of attention). The orientation-specific attention effects, on the other hand, preceded the earliest sensory effect of orientation and some attention effects occurred already around stimulus onset. Although in sensor-space we observed a significant enhancement in neural activations after ∼70 ms poststimulus compared to attention to motion (see Fig. 2 B), which is in line with the sensory orientation effect, we showed that the connectivity dynamics for attention to orientation tend to precede stimulus onset. In right parietal and occipital areas, we found increased local efficiency for attention to motion than attention to orientation already before stimulus onset ( − 28-0 ms), in the alpha and beta-band (see Fig. 3 ). These differences were due to significant decreases in the local efficiency of those ROIs for attention to orientation, already ∼0-100 ms (see Fig. 4 B). We furthermore showed that the pattern of alpha-band connections disengagement started at ∼20-50 ms when attending orientation (see Fig. 6 A). This may occur because orientation is processed earlier (see Fig. 7 B), possibly abetted by the limited variability in the inter-trial interval (see also Section 4.6 ).

Right-lateralization of FBA effects
We found significant ERP differences due to FBA over the right occipito-temporal and central occipital groups of electrodes (see Fig. 2 ). This right-lateralization of attention-induced effects was confirmed also in source-space analyses, where we observed a network reorganization (see Fig. 3 ) and connectivity changes (see Fig. 5 ) involving predominantly areas in the right parietal and occipital cortices. We localized the cortical sources of the sensory motion effect close to the motion-sensitive areas V5/hMT, in the right hemisphere (see Fig. 8 A). Together our findings support the notion of a right hemispheric advantage, which has been previously reported for task conditions involving coherent motion direction in RDKs ( Niedeggen and Wist, 1999 ;Patzwahl et al., 1994 ), and localizers for motion processing ( Plomp et al., 2010 ). For the attentional function, the notion of right hemispheric dominance derives from lesion studies of patients with neglect syndrome, which proposed that the right hemisphere directs attention to both visual hemi-fields, while the left hemisphere directs attention only to the right hemi-field ( Heilman and Van Den Abell, 1980 ;Mesulam, 1981 ). An account for such asymmetric dominance in attentional control was provided by neuroimaging studies, which showed asymmetries in the strengths of attentional signals ( Szczepanski et al., 2010 ). Neuroimaging studies also recognized a right hemispheric dominance for the ventral attention network, which is predominantly associated with stimulus-driven control and responds along with the dorsal (top-down) attention network when behaviorally relevant visual information is detected ( Corbetta and Shulman, 2002 ;Corbetta et al., 2008 ). Our present findings support and extend these previous studies on the notion of right hemispheric dominance for attention, by revealing frequency-specificity (see Section 4.1 ) and timings (see Section 4.4 ) of the attentional biasing signals underlying feature selection.

Limitations and future work
The present study has some limitations and possible extensions that could be examined in future works. The large number of trials contributed to guarantee robust analyses despite we used a limited number of subjects in the two experiments ( N = 20 in Experiment 1; N = 18 in Experiment 2). Nonetheless, considering that all subjects were healthy young adults, it is not straightforward that our results will generalize to other subjects. A possibility for future investigations could be to attempt a replication of our findings with other subjects, and eventually utilize also groups of subjects with attention deficits, such that distinct physiological and pathological patterns of release from gating could be identified.
In this study, we adopted a block design, by using the same task condition for all trials within a block, and we randomly varied fixation interval for every trial, in order to avoid confusing FBA-induced effects with those deriving from mechanisms of trial-to-trial attentional reorienting towards one feature or the other. Nonetheless, subjects possibly defocused their attention after having responded, until just before the next stimulus onset, even between trials with same attentional demand. Furthermore, subjects might have developed an implicit sense of the duration of the inter-trial (fixation) interval, and also of the variability of this interval. For these reasons, we cannot entirely exclude the possible confound of anticipatory mechanisms of attentional defocusing and reorienting towards the same relevant feature. Future work should consider a modified experimental design to try to disambiguate the dynamics of these anticipatory mechanisms.

Conclusion
To conclude, we show FBA to be an inherently networked process that is characterized by quick and precise dynamics. Our findings provide the first evidence that FBA is controlled by a fast, dynamic reorganization in the topology of a functional network of cortical areas. Furthermore, they show that this reorganization is supported by selective decreases of functional influences in the alpha and beta-band, with precise and selective timings. These rapid connectivity changes support the unfolding of perceptual processing according to the current goals, enhancing the local processing and information routing of behaviorally relevant signals. The present findings not only advance our understanding of the timing of FBA and the underlying regulatory mechanisms but are informative for other existing models of attention, particularly those operating on the basis of holistic objects and more complex stimuli like faces and houses (object-based attention). These object representations are known to be associated with specialized areas in the inferior temporal-occipital cortex, like fusiform face area and parahippocampal place area ( Epstein and Kanwisher, 1998 ;Kanwisher et al., 1997 ;O'Craven et al., 1999 ; for review see Scolari et al., 2014 ); we thus expect their selection to be mediated by stereotyped patterns of release from parietal gating, similar to those observed here. Future work could consider using the same methodological framework adopted here, to confirm that specific patterns of dynamic release from gating underlie also FBA on other features (e.g., using color and motion as visual features) ( Schoenfeld et al., 2014( Schoenfeld et al., , 2007, object-based attention (e.g., using faces and house) ( Baldauf and Desimone, 2014 ), and attention to different sensory modalities (e.g., using auditory and visual stimuli) ( Lakatos et al., 2008 ).

Credit author statement
All authors conceived and designed the study. M.F.P. and G.P. coordinated the study. M.F.P. collected and analyzed the data. M.F.P. and D.P. implemented the methods. G.P. provided funding. M.F.P. wrote the first draft of the manuscript. All authors discussed the results, commented on the manuscript, and reviewed it.

Data availability
All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary Materials. Both sensorspace and source-reconstructed EEG data, together with individual lead field matrices, are available on the Open Science Framework ( https://osf.io/c4gv6/ ). Additional data related to this paper may be requested from the corresponding author.