Prestimulus feedback connectivity biases the content of visual experiences

Significance Ongoing neural activity influences stimulus detection—that is, whether or not an object is seen. Here, we uncover how it could influence the content of what is seen. In ambiguous situations, for instance, ongoing neural fluctuations might bias perception toward one or the other interpretation. Indeed, we show increased information flow from category-selective brain regions (here, the fusiform face area [FFA]) to the primary visual cortex before participants subsequently report seeing faces rather than a vase in the Rubin face/vase illusion. Our results identify a neural connectivity pathway that biases future perception and helps determine mental content.

with Matlab (1) and the Psychophysics Toolbox (2), and corrected the timings using a photo diode. The procedure of the experiment is illustrated in the main text Fig. 4.

Behavioural Analysis
We collected behavioural reports after the end of each trial, giving us 400 responses. To test for the stochastic nature of the response we used curve-fitting procedures from the curve-fitting toolbox in Matlab. Specifically, for each participant we binned the data according to how many trials in a row they responded with the same perceptual report. We broke this down in 11 bins with 0 repetitions to 10 repetitions, averaged the number of repetitions within each bin across participants, and then fit the averaged data to a binomial distribution generated in Matlab across the 11 bins before calculating goodness-of-fit.

MEG Data Acquisition
We carried out the MEG recordings using a 306-channel whole-head VectorView MEG system (Elekta-Neuromag, Ltd., Helsinki, Finland, 204 gradio-and 102 magnetometers) installed in a magnetically shielded chamber (AK3b, Vakuumschmelze Hanau, Germany), with signals recorded at 1000 Hz sample rate. Hardware filters were adjusted to band-pass the MEG signal in the frequency range of 0.01 Hz to 330 Hz. Prior to the recording, we recorded points on the participant's head using a digitizer (Polhemus, VT, USA). These points included the 5 HPI coils, the three fiducials (nasion, left and right pre-auricular points), and over 200 additional points covering the head as evenly as possible. We used the HPI coils to monitor head position during the experiment.

MEG Preprocessing and Source Projection
We pre-processed the data using the Fieldtrip toolbox (3). From the raw continuous data, we extracted epochs of 4 seconds lasting from 2.5 seconds before onset of the picture to 1.5 seconds after onset of the picture. This resulted in 400 trials per participant. We applied a highpass filter on this epoched data at 1 Hz (IIR Butterworth 6-order two-pass filter with 36 dB/oct roll-off), followed by a band-stop filter of 49 -51Hz to remove power line noise. We then downsampled the data to 400 Hz. We visually inspected the trials for strong non-physiological artefacts (e.g. channel jumps) and rejected the contaminated trials before computing ICA. We removed components representing typical physiological artefacts (e.g. blinks , ECG) and reconstructed the cleaned raw data. We finally removed the remaining noisy trials by visual inspection. For each participant, we then assigned the trials to the 2 conditions according to the participants' response. Although there was almost equal incidence of face and vase reports to start with, the balance of trial numbers changed after artefact rejection. To ensure a similar signal-to-noise-ratio across conditions, we equalized the trial numbers of face and vase reports by random omission (percentage of trials left in the analysis: M = 79.32%, SD = 15.12%).
We projected the data to source space by applying LCMV (linear constrained mean variance) beamformer filters to the sensor level data (4). We created anatomically realistic headmodels (5) using participants' individual structural MRI and the Polhemus digitized scalp shape. For three participants for which we could not obtain an individual MRI, we used a template MRI which was morphed to fit the individuals head shape using an affine transformation. We calculated a three-dimensional source grid (resolution: 8 mm) covering an entire MNI standard brain volume. For each of these points, we computed an LCMV filter using the individual leadfield and the data covariance matrix (estimated separately for the focus of analysis; see below). We used these spatial filters to then project classifier weights into source space and compute oscillatory and connectivity measures for distinct ROIs (see next sections).

Multivariate Pattern Analysis (MVPA)
We resampled the MEG data to 100 Hz to speed up the MV PA computations (6, 7) using an algorithm that first applies a low-pass filter at one third of the resampling frequency. So, we performed the decoding analysis on the broadband 1-33 Hz time-domain signal. We used MNE Python (8) which uses Scikit-learn (9, 10) for the decoding and implemented a 4-fold crossvalidation procedure within each subject. The analysis was shifted over time on a sample -bysample basis. For each time-point at each sensor, we Z-normalized the MEG data, trained a Logistic Regression classifier on three folds, and tested on the left-out fold. We operationalized the decoding performance as Area Under the Curve (AUC).
To find out which brain regions contributed to above chance decoding performance the most, we used the classifier weights that the classifier used to separate face from vase reports. To obtain interpretable sensor-level topographies, we multiplied the classifier weights by the data covariance in a first step (11). Then we applied LCMV beamformer filters (using a -.3 to .35 s window to calculate the covariance matrix) to project the weights into source space. At the source level, we abolished polarity differences across voxels by taking the absolute values. This approach is near-identical to the "informative activity" procedure reported in a recent study (12). Finally, we averaged the source-level weights across the intervals 50 to 120 ms and 120 to 200 ms and applied a 95%-max threshold to mask our ROIs. This resulted in a V1 ROI centred around MNI coordinates [12 -88 0] mm with a size of 32 grid points and a FFA ROI centred around MNI coordinates [28 -64 -4] mm with a size of 1 grid point (8mm grid resolution).

Analysis of Post-stimulus Oscillatory Activity
The MVPA provided a clear ROI in right ventromedial temporal cortex corresponding to Fusiform Face Area (FFA) at expected latencies (around 160 ms). We performed time -frequency analysis specifically for this ROI by using the single-trial source-projected time series (using the full epoch length for calculation of the covariance matrix). We estimated power using multitaper Fast Fourier Transform (FFT) with discrete prolate spheroidal sequences (dpss) (13), with separate window lengths (.5 s for low frequencies [2-30 Hz in 1 Hz steps] and .3 s for high frequencies [33-99 Hz in 3 Hz steps]). We adapted smoothing to the specific frequencies for which we estimated power (linearly increasing from 2 to 6 Hz for low frequencies and set to +/-20% of the center frequencies for high frequencies). As a control analysis, we repeated the same procedure for high frequencies using window lengths of 0.2 s and 0.1 s.

Analysis of Pre-stimulus Power, Coherence, and Granger Causality
In addition to the face-sensitive region FFA, the classifier weights source analysis implied the involvement of V1 at earlier time points. We calculated power, coherence, imaginary part of coherency, and Nonparametric Granger causality (14) in the pre-stimulus period between FFA and V1 in source space (using the full epoch length to calculate the covariance matrix). We used multi-taper frequency transformation with a spectral smoothing of 2 Hz to get Fourier coefficients in the pre-stimulus period (-1 to 0 s), after which we extracted power and computed coherence and bivariate Granger causality. This gave us separate estimates of connection strengths from FFA to V1 ("feedback") and vice versa ("feedforward"). We repeated the same Granger causality analysis on time-reversed data, expecting reversals in the directionalities of the estimates to rule out spurious connectivity results (15). That is, we expected the feedforward Granger estimates of the original data to differ from those of the time -reversed data, and to instead resemble the feedback estimates of the original data, and vice versa.
We averaged all results over all grid points within the V1 ROI.
As a control analysis, we calculated time-and frequency-resolved coherence based on the timefrequency data obtained as described in the previous section, and averaged across the frequencies between 50 and 90 Hz to obtain a coherence time-course.

Statistical Analysis of MEG Data
For the MVPA analysis, we tested decoding performance against chance level (50%) using a dependent-samples T-test. Since we were interested in periods in which the classifier performs above chance, we used a one-sided T-test. For all remaining statistical analyses, we used nonparametric cluster permutation tests (16), comparing a selected test statistic against a distribution obtained from 10000 permutations. We set thresholds for forming clusters at p < .05 and considered an effect significant if its probability with respect to the nonparametric distribution was p < .05. For the post-stimulus time-frequency contrast in FFA, we tested power in the face-vase contrast in the window of 0 to .35 s, separately for the low and high frequencies, using 2-sided dependent-samples T-tests, and without applying a baseline correction. We did the same for pre-stimulus power, except that this estimate was not timeresolved. We tested coherence and feedforward and feedback connectivity with 1-sided dependent-samples T-tests as we had hypothesized greater values of these measures on face trials compared to vase trials. We restricted the statistical testing window of coherence and Granger to the frequency window 5-25 Hz, which covers the visible peaks in the grand-averaged power and coherence spectra.

Figure S6
: Pre-stimulus contrast (face vs vase) of imaginary part of coherency produces the same result as the coherence contrast, but with less high frequency noise. Red line is imaginary coherency on face trials and blue line on vase trials. Shaded error regions represent the standard error of the mean for within-subject designs (17). Compared to vase trials, face trials show increased pre-stimulus imaginary coherency between V1 and FF1 in the alpha/beta frequency range. Figure S7: Granger Causality estimates on time-reversed data reveal an expected reversal in the directionality of results compared to original data, thereby increasing confidence that the observed results are not a product of co-varying noisy sources. Shaded error regions represent the standard error of the mean for within-subject designs (17). A) Feedforward (V1 -> FFA) Granger causality estimates on original data (a copy of Fig. 2C, left). B) Feedback (FFA -> V1) Granger causality estimates on original data (a copy of Fig. 2C, right). C) V1 -> FFA Granger Causality estimates on time-reversed data resemble FFA -> V1 in original data (B), but not V1 -> FFA in original data (A). D) FFA -> V1 Granger Causality estimates on time-reversed data resemble V1 -> FFA connectivity in original data (A), but not FFA -> V1 in original data (B).

Figure S8:
No correlation between feedback Granger Causality and percent face reports across participants. r value represents Pearson's correlation coefficients. Shaded area represents 95% confidence intervals.

Figure S9:
No correlation between the maximum gamma effect (obtained from the 100 ms analysis window data) and either pre-stimulus feedback Granger or post-stimulus decoding accuracy. r value represents Pearson's correlation coefficient. Shaded area represents 95% confidence intervals.