Willed Attentional Selection of Visual Features: An EEG Study

Visual selective attention studies generally tend to apply cuing paradigms to instructively direct observers’ attention to certain locations, features or objects. However, in real situations, attention in humans often flows spontaneously without any specific instructions. Recently, a concept named “willed attention” was raised in visuospatial attention, in which participants are free to make volitional attention decisions. Several ERP components during willed attention were found, along with a perspective that ongoing alpha activity may bias the subsequent attentional choice. However, it remains unclear whether similar neural mechanisms exist in feature- or object-based willed attention. Here, we included choice cues and instruct cues in a feature-based selective attention paradigm, allowing participants to freely choose or to be instructed to attend a color for the subsequent target detection task. Pre-cue ongoing alpha oscillations, cue-evoked potentials and target-related steady-state visual evoked potentials (SSVEPs) were simultaneously measured as markers of attentional processing. As expected, SSVEP responses were similarly modulated by attention between choice and instruct cue trials. Similar to the case of spatial attention, a willed-attention component (Willed Attention Component, WAC) was isolated during the cue-related choice period by comparing choice and instruct cues. However, pre-cue ongoing alpha oscillations did not predict the color choice (yellow vs blue), as indicated by the chance level decoding accuracy (50%). Overall, our results revealed both similarities and differences between spatial and feature-based willed attention, and thus extended the understanding toward the neural mechanisms of volitional attention.


I. INTRODUCTION
H UMAN brain receives huge amount of information input at each moment.Due to the limited processing capacity of brain, attention has evolved as a fundamental mechanism that enables brain to selectively process task-relevant information and deploy the perception resources.To explore the neural mechanisms of visual selective attention, previous studies typically applied the attention cuing paradigms [1] in which an external cue or instruction directed observers' attention to target objects, positions or features.The investigations of cued attention contributed to establish the top-down attentional control system, especially in spatial attention [2].
However, this widely applied cuing paradigm describes only a particularly specific condition for the actual attention process in the real world.In most cases, attention is concentrated volitionally without external instructions, which was referred to as "willed attention" by Bengson et al [3], [4], [5], [6].In their study, the instruct cue was replaced by a choice cue that signaled the participants to freely attend left or right in a spatial attention experiment.By contrasting choice trials and instructed trials, two event-related potential (ERP) components over frontal and central scalp regions, i.e., EWAC (Early Willed-Attention Component) and WAC (Willed-Attention Component), were identified as potential neural correlates of willed attention, and each of them uniquely correlated with specific attention-related brain regions (e.g., anterior insula, middle frontal gyrus) measured with BOLD activation using functional magnetic resonance imaging (fMRI) [3].These results suggest that a volitional attentional control system might exist for the guidance of spatial selective attention, while whether such a mechanism can be generalized to other formats of selective attention, i.e., feature-or object-based, still remains unknown.
Compared with instructed attention, one intriguing question specific to willed attention is whether the choice made by observers can be predicted from the brain activities.In another study reported by Bengson et al. [7] with the same willed spatial attention paradigm, it was found that the decision of choice (attend left or attend right) could be predicted by the ongoing occipital alpha oscillations during the −800 to 0 ms interval preceding the choice cue, that is, observers with greater alpha power over the left occipital cortex were more likely to choose to attend left in trials.This observation can be directly associated with the well-documented post-cue alpha lateralization in visual spatial attention.Specifically, when participants are cued to shift their attention to the left (right) visual field, there is a greater suppression of alpha power over the right (left) posterior scalp [8], [9], [10].Such a similarity between pre-cue and post-cue alpha activity thus indicates that voluntary attention decisions could be biased by the pattern of ongoing brain activity, as alpha oscillations were supposed to reflect functional inhibition of task-irrelevant information [11], [12].Considering that alpha activity might play a similar role in feature- [8] and object-based [13] attention, one might be interested in whether the prediction of choice from ongoing alpha activity in spatial willed-attention can be generalized to feature-or object-based attentional task.
Here we sought to answer the above questions in a feature-based attention cuing paradigm by recording the ERPs and ongoing alpha activity when participants spontaneously choose to attend stimulus with one specific color (yellow or blue) in each trial.To minimize the attention shift in space, we presented a superimposed stimulus named random moving dot kinematograms (RDKs) with target and distractor features with the same orientation offset from center position [14], [15], [16], [17], [18].Stimuli dots were flickered at a specific frequency, which enabled us to measure steady-state visual evoked potential (SSVEP) responses to attended/ignored target features as a neural index of attentional processing.We hypothesized that if a similar volitional attentional control system exists in feature-based willed attention, we may extract the ERP components related to willed attention [3] by comparing choice cues with explicit instruct cues.
Furthermore, we would test whether participants' choice of attended features could be predicted through ongoing alpha activity.Unlike spatial attention where the contralateral and ipsilateral alpha activity (with respect to attended visual hemifield) could be represented as a lateralized pattern, much less is known about the behavior of alpha activity in featurebased attention.Thus, we sought to investigate whether the alpha power could show any decodable pattern across the scalp between the choices of attending blue/yellow before the choices were made using multivariate pattern analysis as described in [7], [19], and [20].If such a predictive pattern of alpha activity existed in willed attention to different colors, it might suggest a ubiquitous role of alpha activity across different forms of willed selective attention.On the other hand, considering that the similar lateralized pattern between pre-cue and post-cue alpha activity could be a special case in spatial attention, it was possible that such a predictive role of alpha activity was confined to spatial attention.

II. METHOD A. Participants
Thirty-nine healthy college students (25 female; mean age: 25 years; age range: 20-31 years) were recruited from Shanghai Jiao Tong University as participants in this experiment.In our experiment, we determined the necessary sample size based on the previously reported effect sizes including SSVEP amplitudes (d = 0.5 -0.7) [14], [21], EWAC (η 2 p = 0.244), WAC (η 2 p = 0.356) [3], and visual alpha band activity (d = 0.7) [22].A sample size of 34 was required according to G * Power 3.1 [23] by setting the power (1-β) of 0.8 and the α error probability of 0.05.Five participants were excluded in further analysis (one subject made only 4 correct responses in choice condition; one subject made unbalanced color choices of attending blue dots in only 14 trials; three subjects were excluded due to excessive artifacts in EEG), resulting in a total of 34 participants in the final analysis.A mini-mental state examination (MMSE) was taken prior to the experiment to ensure their basic abilities of language, orientation, attention and calculation.Each participant signed an informed consent before the experiment, and received financial compensation for their participation.The study was designed and conducted according to the Declaration of Helsinki, and was approved by the institutional review board of Shanghai Jiao Tong University.

B. Stimuli
Stimuli was created by Psychophysics Toolbox Version 3 (PTB-3) implemented in MATLAB R2018b and displayed on a 23.8-inch monitor (Skyworth: F24G1V) with a resolution of 1920 × 1080 and a refresh rate of 120 Hz.The monitor was placed 60 cm away from participant.Stimulus consisted of two circular random dots kinematograms (RDKs) which were presented centrally in two different colors (yellow and blue) with a 5.5 • radius of visual angle and made random Brownian motion at 0.03 • per frame.Each RDK was composed of 128 overlapped dots with a size at 0.4 • flickering at a specific frequency (blue dots at 15 Hz and yellow dots at 20 Hz, iso-luminant 250 cd/m 2 ), with the background set to dark grey (5 cd/m 2 ).Flicker frequencies were selected outside alpha-band range to avoid interference to potential modulation of alpha-band activity.The RDKs lasted for 2000 milliseconds per trial, and coherent motion towards four directions of the Cartesian coordinates arose twice during this latency.Random dots with the same color would move simultaneously towards the same direction during each coherent motion, and two coherent motions within the same trial should be separated at least by 800 ms.Coherent motion never began within the first 300 ms after the onset of RDKs, and always ended 300 ms before the offset of RDKs.Color of RDKs that showed coherent motion was randomly decided each time [24].Only 75 percent of random dots joined the coherent motion each time to avoid participants gazing at one specific dot.

C. Main Task
The experiment mainly included two within-subject conditions.The main difference of the two conditions is that participants would be "instructed" to direct attention in the instruct cue trial and would "choose" to attend whichever target they wanted in the choice trials.In the beginning of the instruct trial, a colored-rhombus cue was presented together with the fixation cross (Fig. 1).Participants should focus on the central fixation crosshair, and covertly shift attention to the random dots with the color (blue/yellow) filled in the rhombus cue in the instruct cue trials.Following a jittered duration of 1000 -1200 ms, RDKs were presented for 2000 ms.Participants should press the up arrow key as fast and accurately as possible whenever they detect the coherent motion (brief duration 300 ms) of the to-be-attend color RDKs but ignore the movements for uncued color.On the other hand, in the choice trials, participants should choose the color to attend on their own when neither of the two colors was filled in the rhombus cue.To avoid participants adopting pre-defined strategies to determine the color to be attended (i.e.always choose the same color, or exchange color choice in each trial), we asked them to randomly choose a color to attend immediately once they detect the rhombus without color cue.The motion-detect task was the same as that in the instruct cue trials.Participants should press the left or right arrow key to report their choices when the word "Report" appeared at the screen.
In addition, no-target trials were designed as control trials in case that participants responded randomly without distinguishing the target.In no-target trial, there is no rhombus cue nor target (coherent motion) on the screen, participants should focus on the central fixation without making any reaction on the keyboard.Overall, the experiment included 504 trials, which consisted 238 instruct cue trials, 238 choice cue trials, and 28 no-target trials.Noted that there are trials with only one occurrence of coherent motion in both instruct and choice trials which constitute 5% of the total trials.Besides, we collected the resting state EEG of eye-closed and eye-open before and after the whole experiment for 3 mins respectively, 12 mins in total.

D. Data Acquisition and Preprocessing
Continuous EEG data was recorded with a 64-channel electrode cap (Easy cap, Brain Product Inc., Germany), with FCz as reference and AFz as ground.An additional electrode was placed below the right eye to monitor ocular activity.
Electrode impedance was kept below 10k for most electrodes, and few electrodes (usually FT7 and FT8) could fail to be kept below 10k but never exceeded a maximum impedance of 50k .The EEG signals were recorded by amplifier (BrainAmp, Brain Product Inc., Germany) with online passband from 0.016 Hz to 250 Hz and sampling rate at 1000Hz EEG data was preprocessed off-line with MATLAB-based EEGLAB (Version: 13.5.4)[25] and ERPLAB (Version: 7.0.0)[26] toolboxes.First, the raw continuous EEG data was filtered into 0.1-40 Hz by a two-way band-pass Butterworth filter with zero phase shift and a roll-off slope of 12 dB/oct to minimize DC drifts and high frequency noises.A notch filter at 50 Hz was further applied to remove remaining power-line inference.EEG data was then down-sampled to 250 Hz.Next, ocular artifacts and remaining noises (i.e., muscle activity) were corrected by independent component analysis (ICA) using the Infomax algorithm [27].Typically, we removed one component related to eye movements and one component related to eye blinks for each subject.Artifacts-corrected EEG data were then re-referenced to the average of 2 mastoid electrodes (TP9 an TP10), and the original recording reference electrode FCz was recovered at the same time.After that, continuous EEG data recording reference electrode FCz was recovered at the same time.After that, continuous EEG data were segmented into two types of epochs, i.e., one time-locked to cue onset (−1500 to 1000 ms) and the other time-locked to target onset (−500 to 2000 ms).It should be noted that the epochs were segmented a little longer than the periods we were interested in, so that we could trim the two ends of the epoch (i.e., 500 ms) to minimize the edge artifacts caused by filtering in the following EEG analysis.Epochs were baseline corrected by subtracting the average of pre-cue/pre-stimulus voltages.EEG epochs were excluded from further analysis if (1) any point containing voltages exceeding ± 100µV or (2) epochs containing peak to peak activity that is greater than 100µV within a moving window (window length: 1000 ms), as implemented in pop_artmwppth.m function ERPLAB).On average, the percentage of excluded trials with artifacts was 9.10 % ± 6.19 % (mean ± SD) across subjects.

E. Data Analysis
1) Behavioral Analysis: Only the key presses that occurred between 200 ms and 900 ms after the onset of a coherent motion were considered as correct responses and included in the analysis of reaction time.The upper boundary was chosen based on the results of pilot tests of the experimental paradigm in which most correct responses were made within 900 ms.Moreover, subjects were not likely to detect the coherent motion and make a correct response within 200 ms.Reaction time was averaged across those correct responses for each coherent motion.Behavioral analysis was performed for each trial type (instruct/choice cue, attending yellow/blue) separately.Two-way repeated-measures ANOVA with cue type (instruct, choice) and attending color (yellow, blue) as within-subject factors was conducted on both accuracy and RTs.
Given that most trials contained two occurrences of coherent motion, we divided these trials into 3 categories based on accuracy: double correct trials (participant made two correct reactions in a single trial), single correct trials (participant made only one correct reaction in a single trial), and wrong trials (participant made no correct reaction in a single trial).On average, double correct, single correct and wrong trials accounted for 67.57% ± 11.02%, 25.46% ± 4.82%, and 6.96% ± 4.30% of total trials, respectively.We selected the epochs with double correct responses in all of the EEG analysis.
2) SSVEP Analysis: Data of the first 400 ms after RDKs onset was discarded, as previous research had suggested that 400 ms should be long enough for RDKs to elicit a stable SSVEP response [28].The selected EEG epochs were transformed to frequency domain using Fast Fourier Transform (FFT).We performed SSVEP power analysis by computing signal-to-noise ratio (SNR) as below: where P denotes SSVEP power spectral estimates, f denotes center frequency and P ( f ) denotes the average power spectral in the frequency bands f − 2, f − 0.5 ∪ f + 0.5, f + 2 [29].
Converting power spectrum to SNR help facilitate comparability across individuals and reduce the effects of 1/f power spectral shifts [30].Electrode with the maximum SNR at the flicker frequency (i.e.,15 Hz, 20 Hz) was chosen as the channel of interest for each participant, which was in line with the functional localizer approach [31] (channel selection for each participant was shown in Fig. 2).For the statistical analysis of SSVEP, SNR of the target frequency was subjected to a three-way repeated measure ANOVA with Cue Type (instruct cue, choice cue), Color (yellow, blue) and Attention (attend, unattend) as within-subject factors.
3) Multivariate Pattern Classification/Decoding of Cue-Related ERP: Cue-related ERPs were obtained by averaging cue-related EEG epochs (−200 to 1000 ms) for each cue type (instruct cue, choice cue), electrode and participant, with the pre-cue 200 ms interval as baseline.According to previous research [3] we expected to observe two ERP components that associated with willed attention: Early Willed Attentional Component (EWAC) occurring 300 to 400 ms post-cue, and Willed Attentional Component (WAC) occurring between 400 and 800 ms after cue onset.These two components were isolated by comparing the cue-evoked responses of choice trials and instruct trials.In this study, we treated multi-channel EEGs as a topographic pattern to perform classification instead of treating them singly in univariate EEG/ERP approach.This method is more sensitive and offers improved statistical power compared to the traditional univariate approach [32].Cue-related ERP waveforms can be found in supplemental materials Figure S4.The decoding procedure and scripts were adapted from a routine to decode visual spatial attention from scalp ERPs [33].Decoding was performed independently at each time point within the epochs of interest (−200 to 1000 ms) using the support vector machine (SVM) classifier trained through the MATLAB fitcsvm() function.To ensure our decoding was free from high frequency interference, we applied a low-frequency filter at 20 Hz using eegfilt() implemented in EEGLAB.Time-resolved EEG data was then down-sampled at 50 Hz (1 data point every 20 ms) in order to reduce the computation cost.Firstly, we reorganized the trials on the basis of cue type (instruct cue, choice cue).Then a 3 fold cross-validation was applied, that is, for each subject, the epochs of each type were randomly and equally divided into three folds, with two folds serving as the training set and the remaining fold serving as the testing set.The final one or two trials would be discarded if total trial number cannot be evenly divisible by 3. Cue-evoked responses were averaged across trials in each group.Moreover, we performed a z-score normalization across channels at each time point to avoid the For each participant, one electrode with maximum SNR was chosen as channel-of-interest.For each channel, the number of subjects for that channel was selected as channel-of-interest was calculated, and further illustrated in the topographic map.
unrelated decodable voltage difference caused by the overall power difference of the two conditions being compared [34].The trained SVM classifier was then used to predict the cue type in the testing set using the MATLAB function predict(), which output a series of prediction of instruct/choice cues.Decoding was considered correct when the classifier correctly determined the type of the cue (instruct cue or choice cue), so the chance performance was 50% for balanced data.This procedure was repeated three times, each time with each of the three folds of EEG epochs serving as the testing set.The above procedure was iterated for 20 times, each time with a new random assignment of trials into the three folds.Such a decoding procedure included 120 decoding attempts (2 trial types × 3 cross-validations × 20 iterations) at each time point, across which an average decoding accuracy was obtained.
If the decoding of cue-related ERPs can reliably classify which of the two cues (instruct cues, choice cues) is perceived, then the decoding accuracy should be greater than chance level.To confirm the conclusion, decoding accuracy was compared against the chance performance using the sign-rank test at each time point.To correct for multiple and comparison, threshold-free cluster enhancement (TFCE) approach was applied with a cluster threshold set as p = 0.05 [35], [36].
If a time cluster with significant decoding accuracy was detected, a permutation test was applied to determine whether this significance was decoded by chance.This approach was similar to Luck et al. [37].For each subject, labels of instruct and choice trials were randomly shuffled and used in the following training, testing, decoding and statistical process as described above.We recorded the maximum length of each cluster with significantly larger decoding accuracy than chance level (if there is no significant cluster was found, it will be recorded as 0).This procedure was repeated 1000 times to construct a null distribution, which represented the maximum length of the mass cluster with significant decoding accuracy when there is no information about the true cue types.A time cluster was considered significant and not generated by chance if its length was on the top 95% of the null distribution.
In addition to decoding accuracy, we've also evaluated the contribution of each channel to drive the performance of classifier.This was achieved by reconstructing the weight map of the SVM classifier at each time point.The weight map is obtained by multiplying the classifier weights with the covariance matrix of the original data [38].

4) Alpha-Band Power Analysis and Decoding Time-Resolved
Alpha Power Pattern: In this study, we were interested in whether pre-cue alpha power could predict participants' choice of attended color.The individual alpha frequency (IAF) was determined by calculating the center of gravity within the alpha band (8-13 Hz) [39] using the average of power spectra within the occipital cluster (i.e., O1, O2, P7, P8, PO3, PO4, P5, P6, PO7, PO8, POz, and Oz) during the pre-task eyeclosed resting EEG.Data was first FFT-transformed and power spectral was averaged across the selected epochs of the EEG channels in the occipital cluster.Then IAF was calculated according to Eq. ( 2) Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

TABLE I BEHAVIOR RESULTS (MEAN±SD) AND RESULTS OF 2-WAY REPEATED MEASURE ANOVA IN DIFFERENT TRIAL TYPES
in which A( f ) the power spectral estimates at frequency f , f 1 and 2 are chosen as 8 Hz and 13 Hz in the alpha range.
The pre-cue alpha power for choice blue and choice yellow was calculated using the individual alpha frequency (IAF) of each participant.EEG data was extracted in the time window of interest, i.e., from 800 ms to 0 ms before cue onset.For each participant, the time-domain EEG epochs were transformed into frequency domain by means of FFT, and further averaged across trials of the same condition (choice, instruct).Then the power was averaged within the individual alpha band defined as [IAF-2, IAF+2] Hz for each electrode and trial condition.Results were displayed on the topographic map Fig. 4a-b.
To quantitatively determine whether pre-cue alpha band activity pattern is predictive to the subsequent color choice (choose blue vs. choose yellow) in feature-based attention, we performed a multivariate decoding analysis on alpha power during the pre-cue period.Data of choice trials were extracted from 1500 ms pre-cue to cue onset, and were baseline corrected to the average of 1500 ms to 1000 ms pre-cue.To obtain the time course of pre-cue alpha-band power, we applied a band-pass filter with the cut-off frequency at [IAF-2 Hz, IAF+2 Hz] using eegfilt() implemented in EEGLAB, and then calculated the centered moving average at a window size of 100 ms.Procedures of decoding and statistical analysis of the decoding accuracy were consistent with cue-related ERP analysis which were described above.

A. Behavioral Results
The overall accuracy across all trial types was 80.95% (SD = 6.06%), with average hit rate of 78.56% (SD = 11.91%)average false alarm rate of 23.18% (SD = 12.90%).Average reaction time was 720.61 ms with SD = 25.92 ms.These parameters were separately analyzed for different types of trials, as described in Table I.Two-way repeated-measures ANOVA on accuracy revealed significant main effects of Attending Color.No significant main effect of Cue Type nor significant Attending Color × Cue Type interaction was found.Similarly, the ANOVA on RTs showed a significant main effect of Attending Color, indicating that participants responded faster to yellow targets than to blue targets.No significant main effect of Cue Type nor significant Attending Color × Cue Type interaction was found.In summary, these results suggest an overall moderate level of task difficulty (∼80% accuracy on average), and that participants had better performance (higher accuracy and faster responses) when attending yellow dots than attending blue ones.

B. Attentional Effects Represented by SSVEP Power Spectrum
Topographic distributions of the electrode selected for each participant at both target frequencies (i.e., 15 Hz, 20 Hz) were presented in Figs.2c and 2d.Grand-averaged power spectrum of SSVEP converted into SNR units was depicted in Figs.2a  and 2b for instruct and choice cue condition respectively.
The power spectrum showed two decisive peaks at 15 Hz (evoked by blue flickering dots) and 20 Hz (evoked by yellow flickering dots), and the SNR was generally greater when the corresponding color was attended than it was ignored, as indicated by the significant main effect of Attention (F (1,33) = 28.679,p < 0.001, η 2 p = 0.503) observed in the three-way ANOVA on SNR.There is also significant Attention × Color interaction effect (F (1,33) = 4.400, p = 0.044, η 2 p = 0.118).Paired-samples t-tests suggest that attention effects were significant for all combinations of Cue Type and Color (all ps < 0.001).There was no main effect for either Color (F (1,33) = 1.873, p = 0.180, η 2 p = 0.054) or Cue Type (F (1,33) = 1.689, p = 0.203, η 2 p = 0.049).To further parse the Attention × Color interaction, we computed the SNR ratio (attended vs. unattended) as attentional modulation for all combinations of Cue Type and Color, and tested the SNR ration with a two-way repeated measure ANOVA with Cue Type and Color as within-subject factors.Results of this follow-up ANOVA for attentional modulation showed a main effect of Color (F (1,33) = 52.168,p < 0.001, η 2 p = 0.613), suggesting a significantly greater attentional modulation of 15 Hz (tagged on blue dots) SSVEPs than that of 20 Hz (tagged on yellow dots) SSVEPs, which explains the Attention × Color interaction in the initial three-way ANOVA.The main effect of Cue Type (F (1,33) = 1.962, p = 0.171, η 2 p = 0.056) and the interaction of Cue Type × Color (F (1,33) = 0.027, p = 0.869, η 2 p = 0.001) in the follow-up ANOVA were not significant.In summary, these results suggest that feature selective attention could modulate the magnitude of SSVEP responses no matter attention was deployed by an instructive cue or the spontaneous choice of the participant.Considering that only trials with double correct responses were included in the analysis of SSVEP, different trial numbers between yellow and blue targets due to accuracy differences might be an underlying impact on SSVEP responses.To test the effect of trial numbers, we equalized the two types of targets by randomly selecting part of 'attend yellow' trials that had higher accuracy for all participants and conducted the same three-way repeated measure ANOVA.Results indicated a significant main effect of Attention (F (1,33) = 32.202,p < 0.001, η 2 p = 0.494) and no significant main effect of Color (F(1,33) = 0.984, p = 0.328, η 2 p = 0.029), which was similar to that when including all trials for analysis.Therefore, our SSVEP results were not likely biased by the differences in trial numbers.

C. Decoding Accuracy of Instruct / Willed Attention in Cue-Related ERPs
Figure 3e demonstrated that decoding accuracy generally exceeded the chance level following the cue onset.Three time-clusters exhibited higher decoding accuracy than the chance level, as confirmed by the initial sign rank test.These clusters occurred respectively within the intervals [200,240] ms, [460, 480] ms, and 980] ms.The timing of these clusters aligned with the occurrences of EWAC (300 to 400 ms post cue) and WAC (400 to 800 ms post cue) mentioned by previous study [3].However, only the cluster of [680, 980] ms passed the permutation test (the upper 5 th percentile = 100 ms).The weight map showed in Figs.3a-d

D. Decoding Accuracy of Pre-Cue Alpha Activity During Willed Color Attention
We investigated whether pre-cue alpha power contained information that could predict participant's choice of willed color attention based on the alpha power pattern over the scalp, as previously found in willed spatial attention [7].Figures.4a, 4b and 4c illustrated the topographic maps of alpha power averaged between −800 ms and 0 ms before cue onset for choice yellow, choice blue and their differences, Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
respectively.IAF of all participants were in the range of [9.29, 11.11] Hz, with the median IAF equal to 10.06 Hz.An SVM model was trained and tested using pre-cue alpha power to determine whether a differentiable pattern can be decoded between choice blue and choice yellow trials at the scalp level.As shown in Fig. 4d, there was a cluster of time points during [−680, −580] ms which showed above chance level decoding accuracy in the initial sign-rank test.However, this cluster did not pass the permutation test (the upper 5 th percentile = 140 ms).In addition to IAF, we also calculated the decoding accuracy using pre-cue alpha power (8 -13 Hz), which showed no significant decoding accuracy from chance level as well.
Although the focus of alpha analysis in the present study was to test whether pre-cue alpha could predict subjects' choice in a willed feature-based attention task, there are some related open questions that might be preliminarily tested with the current dataset.First, one might be interested in the relation between alpha activity after the cue onset and feature-based anticipatory attention.We conducted an additional analysis to preliminarily test whether alpha power could be decoded between yellow choice and blue choice trials after the cue onset.However, as shown in Figure S1 in the supplemental materials, no significant time cluster with above chance level decoding accuracy was observed.
Second, whether alpha power patterns may discern between willed and instructed attention after the cue onset could be another interesting question.We tried to answer this question by performing a time-resolved alpha decoding between willed and instructed trials after the cue onset.However, as shown in Figure S2 in the supplemental materials, no time cluster was observed with decoding accuracy significantly above the chance level.
Third, as we selected trials with double correct responses in all the EEG analysis, we further examined whether pre-cue alpha prediction was related to subjects' performance by performing an additional decoding analysis on pre-cue alpha power based on single correct trials.Similar to the results based on double correct trials, no time cluster with significantly above chance level decoding accuracy was found, as shown in Figure S3 in the supplemental materials.

IV. DISCUSSION
Compared with the classical attention cuing paradigms in which participants' attention was instructed by an external cue, a more common situation in the daily life is that attention is concentrated volitionally by participants' choices.Previous study has reported event-related potentials as possible neural correlates of the choice process when participants freely chose to attend left or right in a visuospatial attention experiment, and that the ongoing alpha activity over occipital cortex might be able to predict the decision of subsequent attentional choice [7].The main objective of this study was to assess whether such ERPs and alpha activities related to the choice were present when attention was directed to features (color).Our findings revealed a significant time cluster of cue-related ERP decoding which was analogous to the WAC in willed spatial attention from Fig. 3.However, we did not find the predictive role of ongoing alpha activity in willed feature selective attention (Fig. 4d).
In the present study, we chose yellow/blue color as the target feature in view that color is a powerful and efficient feature compared with shape and orientation.It was documented that color-selective attention elicited significantly stronger attentional facilitation, as well as faster RTs, compared with orientation [40].To assess the attentional modulation of early visual processing, we adopted a frequency-tagging SSVEP approach by presenting random dots with each target color flickering at different frequency.As expected, we observed significant attentional modulation, that is, stronger SSVEP responses elicited by targets with the attended color than that with the unattended color (Fig. 2), in both instruct cue and choice cue trials, and that the magnitude of attentional modulation did not differ between the instruct cue and choice cue trials.Additionally, SSVEP at 20 Hz has lower attentional modulations than SSVEP at 15 Hz, which may be due to the aperiodic (1/f-like) component of power spectra that shows exponentially decreasing power with increasing frequency [41].Such results confirmed that participants successfully deployed their attention to the corresponding color dots.
It is worth noting that subjects might make plans for the choice before the cue onset, as they could expect the timing of cue presentation due to the steady inter-trial interval (2000 ms) in our experiment [42].Although we could not rule out this possibility, it is less likely to occur because the three type of trials (instruct cue, choice cue and no-target cue) were presented in a random sequence.Nonetheless, future research should consider to randomize the inter-trial interval and minimize possible impacts on the ERP correlates of willed attention.
Meanwhile, one might speculate that in the choice trials, participants might adopt a strategy to wait to decide until the onset of RDKs, since the coherent motion never began within 300 ms.However, it should be challenging for participants to make a choice decision and shift their attention in such a short time (300 ms), and the randomly presented 5% trials with only one occurrence of coherent motion did not encourage such a strategy.Furthermore, if such a strategy were taken, one might expect that participants should be more likely to be slower in response to the first coherent motion compared with the second one due to the extra decision process.However, on average, there were only 40.32% ± 12.33% of choice trials with slower responses to the first coherent motion compared with the second one, which did not support the abovementioned strategy.
We implemented SVM-based multivariate decoding at each time point during the cue-related period (instruct cue vs. choice cue).This analysis yielded three distinct clusters where decoding accuracy exceeded chance levels within specific time intervals (Fig. 3) though only one had passed the permutation test.By comparing these clusters with significant ERP decoding accuracy against the ERP components identified in a study on willed spatial attention [3], we observed overlapping latencies which might suggest potential connections.Notably, the significant cluster spanning [680 ms, 980 ms] corresponded with a significant WAC previously isolated in a feature-based willed attentional task, characterized by an occipital distribution.Additionally, a cluster between [200 ms, 240 ms], which did not reach significance in the permutation test, seemed to exhibit resemblance to EWAC.EWAC has been speculated to be a task-relevant variant of P300 or a P300-related process that suggested a categorization of choice cue and WAC may indicate the subsequent process was related to choose a willful color to attend [43], which jointly constitute the volitional attentional control system.By showing this ERP effect in feature-based willed attention task, our finding consolidated its feasibility as a marker of different formats of willed attention.Such findings are also consistent with the view that feature-based attention might share similar frontoparietal attentional control mechanisms as spatial attention in the top-down attentional cuing paradigms [44].
Contrary to willed spatial attention, our results of timeresolved alpha-band decoding found no significant prediction effect of pre-cue alpha on the choice behavior of attended color, which is against the idea that the pattern of ongoing alpha-band power is predictable for subsequent choice in willed feature-based attention.Although we could not exclude the possibility that the difference of ongoing alpha power pattern preceding yellow vs. blue choices (Fig. 4c) is present but too tiny to be distinguished in the present data, it should be noted that whether alpha activity over the scalp has a decodable pattern between different conditions in feature-based attentional control still remains largely unknown.Focusing on feature-based attention, new evidence was recently found by Muller et al. [45].They claimed that alpha-band activity has no association with global color attentional selection in a top-down cuing paradigm, which is similar to our results.Contrasting with Foxe et al. [46], a possible speculation may be that the alpha-band modulations only work on selection across different feature dimension, but not on different features within a single dimension.On the other hand, alpha activity has been hypothesized as a neural signature of functional inhibition of task-irrelevant information [11], [47], [48], and its lateralization modulation has been closely linked to spatially shifted attention as reported by numerous studies [48], [49].While interestingly, there is a recent revival of research regarding the functional role of alpha activity in spatial selective attention, with either evidences in favor of a causal mechanism of cortical activation/deactivation provided by the alpha neurofeedback training [50], or studies against a causal role which showed the later than expected temporal dynamics of alpha activity compared with that of the attentional modulation of sensory processing revealed by the time course of SSVEP responses [22], [51].This may indicate that alpha activity is rather a consequence of involving in attentional modulation process than a dominator of this visuospatial attention mechanism [52].Thus, further understanding toward the role of ongoing alpha activity in predicting the choice behavior should also be benefited from future studies that aim to reveal the behavior of alpha activity during the classical top-down feature-based attentional control.
Another possible explanation to our negative ongoing alpha prediction result (Fig. 4d) is that participants might not routinely deploy their attention to a non-objective color before the trial onset.In the previous willed spatial attention study by Bengson et al. [7] participants were instructed to keep their fixation on the central crosshair throughout the trial, but not during the inter-trial interval.As the main task of this design was to discriminate the peripherally presented target stimulus, such a task context was likely to bring a spontaneous shift of spatial attention toward the target locations during this interval.Thus, we may speculate a covertly shifted attention to the left/right location during the inter-trial interval which might elicit asymmetrical alpha power distribution and also affect the subsequent direction choice.By contrast, it was unlikely to happen when the target was not presented at a specific position but presented in a specific color.Nevertheless, future studies can test this possibility by controlling spatial attention shift during the pre-cue period, e.g., instructing participants to focus on the central fixation and react to a corresponding task [22].
In conclusion, we studied the neural mechanisms of "willed attentional control" in a color-based selective attention task, focusing on the ERP signatures of allocating willed attention versus cued attention and the predictive role of ongoing alpha activity.We measured SSVEPs elicited by targets in the attended/unattended color to confirm the effective attentional modulation of early sensory processing.Compared with the findings reported in the previous study on willed spatial attention, we found a similar ERP over the occipital region that might indicate the choice-related mental processes in the volitional attention control system; however, we did not observe a predictive role of the ongoing alpha pattern between different color choices.Overall, our results extend our understanding toward whether a unified volitional attentional control mechanism exists across the space-and feature-based selection.The explanation of current findings can benefit future research that aims to reveal the functional roles of alpha activity in feature-based selective attention.

Fig. 1 .
Fig. 1.Trial sequence for two attention tasks of instruct/choice cue.(a) Main task for instruct attention.A rhombus cue on top of the fixation cross indicate the color participant should pay attention to in the following task.Participants covertly attended to the flickering RDKs with the instructed color and press the up button for detecting up to two coherent motion with 300 ms long.(b) A rhombus without any color cue indicated that participants should independently make a choice of color to attend and perform the subsequent coherent motion detecting task.Participants should report their color choices with left or right key when the report cue appeared on the screen.Report cue would sustain for 1 second at most and disappeared as soon as the participant made a response.(c) In no-target trial, there was no rhombus cue nor target (coherent motion) on the screen.

Fig. 2 .
Fig. 2. SSVEP power spectra represented as signal-to-noise ratio (SNR) units at the individually selected channels.(a-b) Grand-averaged SNR of instruct cue trials (a) and choice trials (b).Solid lines indicated attention on blue dots flickering at 15Hz and dashed lines with asterisk mark indicated attention on yellow dots flickering at 20Hz.The grey shading indicated the ±1 SEM.(c-d) Channel-of-interest at the flickering frequency of 15Hz (c) and 20Hz (d).For each participant, one electrode with maximum SNR was chosen as channel-of-interest.For each channel, the number of subjects for that channel was selected as channel-of-interest was calculated, and further illustrated in the topographic map.

Fig. 3 .
Fig. 3. Time-resolved decoding accuracy of cue-related ERPs between instruct and choice trials.(a-d) Weight maps of 240 ms, 480 ms, 700 ms and 900 ms choosing from clusters of time points in which the decoding accuracy was above chance level in the initial sign-rank test, as marked with grey areas in (e).Chance level performance (0.5) was marked by the horizontal dashed line.The blue shading indicated the ± 1 SEM.
displayed a fronto-occipital distribution at approximately 480 ms, which transitioned into an occipital distribution by 700 ms and 900 ms.

Fig. 4 .
Fig. 4. Topographic maps of pre-cue alpha power (IAF±2Hz) and decoding accuracy of the color choice based on pre-cue alpha power.Pre-cue alpha power when participants chose to attend yellow dots (a), blue dots (b), and the difference between yellow choice and blue choice (c) at the interval of −800 ms to 0 ms before cue onset.(d) Timeresolved decoding accuracy of the color choice (yellow vs. blue) based on the pre-cue alpha power.Black dashed line indicated the chance level performance (0.5).Grey areas indicated the cluster of time points in which decoding accuracy was above chance level in the initial signrank test.The blue shading indicated the ± 1 SEM.