Implicit Neurofeedback Training of Feature-Based Attention Promotes Biased Sensory Processing during Integrative Decision-Making

Angela I. Renton; David R. Painter; Jason B. Mattingley

doi:10.1523/JNEUROSCI.0243-21.2021

Abstract

Complex perceptual decisions, in which information must be integrated across multiple sources of evidence, are ubiquitous but are not well understood. Such decisions rely on sensory processing of each individual source of evidence, and are therefore vulnerable to bias if sensory processing resources are disproportionately allocated among visual inputs. To investigate this, we developed an implicit neurofeedback protocol embedded within a complex decision-making task to bias sensory processing in favor of one source of evidence over another. Human participants of both sexes (N = 30) were asked to report the average motion direction across two fields of oriented moving bars. Bars of different orientations flickered at different frequencies, thus inducing steady-state visual evoked potentials. Unbeknownst to participants, neurofeedback was implemented to implicitly reward attention to a specific “trained” orientation (rather than any particular motion direction). As attentional selectivity for this orientation increased, the motion coherence of both fields of bars increased, making the task easier without altering the relative reliability of the two sources of evidence. Critically, these neurofeedback trials were alternated with “test” trials in which motion coherence was not contingent on attentional selectivity, allowing us to assess the training efficacy. The protocol successfully biased sensory processing, resulting in earlier and stronger encoding of the trained evidence source. In turn, this evidence was weighted more heavily in behavioral and neural representations of the integrated average, although the two sources of evidence were always matched in reliability. These results demonstrate how biases in sensory processing can impact integrative decision-making processes.

SIGNIFICANCE STATEMENT Many everyday decisions require active integration of different sources of sensory information, such as deciding when it is safe to cross a road, yet little is known about how the brain prioritizes sensory sources in the service of adaptive behavior, or whether such decisions can be altered through learning. Here we addressed these questions using a novel behavioral protocol that provided observers with real-time feedback of their own brain activity patterns in which sensory processing was implicitly biased toward a subset of the available information. We show that, while such biases are a normal and adaptive mechanism for humans to process complex visual information, they can also contribute to suboptimal decision-making.

Introduction

To interact with a complex and dynamic visual world, humans accumulate samples of noisy evidence to guide perceptual decisions (Ratcliff and Rouder, 1998; Usher and McClelland, 2001). This process is well characterized for decisions based on isolated sources of evidence (Summerfield and Tsetsos, 2015) but is less understood for decisions that require observers to integrate information across multiple sources (Spitzer et al., 2017). Such “integrative” decisions are particularly vulnerable to bias because different sources of evidence can be subject to different degrees of prioritization of sensory processing (Mostert et al., 2015). Indeed, internal biases in the allocation of sensory resources are ubiquitous in visual processing (Appelle, 1972; Yantis, 1993; Theeuwes et al., 2010; Anderson, 2016; Failing and Theeuwes, 2018). Typical behavioral investigations probe such biases by introducing “bottom-up” differences in evidence presentation, manipulating factors, such as stimulus reliability (Fetsch et al., 2013; Zylberberg et al., 2016), probability, or the reward associated with different visual features (Gao et al., 2011; Summerfield and De Lange, 2014). To eliminate these bottom-up sources of bias, we embedded implicit neurofeedback into an integrative decision-making task, unconsciously rewarding observers for preferential attentional allocation to one source of evidence over another. An ideal observer should give equal weight to two overlapping stimuli of identical reliability (Rahnev and Denison, 2018). Here we asked whether neurofeedback can bias (or “adapt”) neural representations in favor of one source of evidence over another, even though this might be statistically suboptimal. This novel approach allowed us to directly assess how differences in the neural processing of concurrent stimuli relate to how they are weighted in integrative decisions.

During perceptual decisions, observers repeatedly sample the available information and stochastically accumulate evidence toward a decision threshold (Usher and McClelland, 2001; Gold and Shadlen, 2007; Ratcliff and McKoon, 2008). Studies of decision-making based on single stimuli have shown that factors, such as the rate of evidence accumulation, or the decision threshold, vary for different decision alternatives (Salzman et al., 1990; Gao et al., 2011; John-Saaltink et al., 2016). There is also evidence to suggest that integrative decisions are biased toward more reliable stimuli (Raposo et al., 2012; Brunton et al., 2013), even when this leads to suboptimal decisions (Rangelov et al., 2020). We set out to investigate the temporal dynamics of evidence accumulation during integrative decision-making by using an implicit neurofeedback protocol to unconsciously reward prioritized processing of one stimulus over another.

Implicit neurofeedback protocols, in which participants are blind to the specific pattern of neural activity manipulated, have been established as a valuable tool for determining causal links between neural activity and behavior (Purcell et al., 2010; Moxon and Foffani, 2015; Sorger et al., 2019). We embedded neurofeedback within our behavioral task as dynamic changes in task difficulty, thus entirely concealing that participants were receiving neurofeedback (deBettencourt et al., 2015). Participants were asked to report the average direction of motion across two overlapping sets of moving bars, which were tilted at −45° and 45° (Fig. 1d). The two sets of bars moved in different directions, and each set flickered at a different frequency. This flicker induced steady-state visual evoked potentials (SSVEPs), an established index of feature-based visual attention (i.e., prioritized processing of relevant visual features) (Müller et al., 2006; Andersen et al., 2011; Painter et al., 2014; Renton et al., 2019b). Using normalized subtractions of SSVEP amplitudes, we altered task difficulty in real time. Specifically, motion coherence across all bars increased when attention was allocated to bars of the trained orientation, and decreased when attention was allocated to bars of the untrained orientation. To assess carryover effects of neurofeedback, we included test trials in which motion coherence was held constant (i.e., no real-time neurofeedback); thus, the display was no longer contingent on participants' attention. We used a combination of time-frequency analyses and computational modeling to assess how neurofeedback impacted integrative decision-making processes.

Figure 1.

Motion averaging task. a-c, Polar histograms of the number of bars moving in each direction during an example trial. The transparent blue and green histograms represent bars oriented at −45.00° (tilted left) and 45.00° (tilted right), respectively. In this example, left-tilted bars had an average motion direction of 0° (blue arrow), whereas right-tilted bars had an average motion direction of 67.50° (yellow arrow). The grand-average direction of motion was therefore 33.75° (gray arrow). The averaging task was easier when motion coherence was (a) relatively high as opposed to (b) relatively low. c, During the first 1 s of each trial, the averaging task was impossible since movement direction was uniformly distributed for both left- and right-tilted bars. d, Timeline of an example trial. Coherent motion was presented for 2 s following the 1 s period of random motion. On neurofeedback trials, the level of motion coherence of both sets of bars was contingent on SSVEP selectivity for bars with the trained orientation. On test trials, motion coherence was constant for the entire 2 s period. After 3 s of bar motion, participants were cued to indicate the perceived average direction of motion for the trial. The response cursor did not appear until participants moved the mouse. After the participant left clicked the mouse to input a response, a red bar (shown white in figure) appeared for 1 s, displaying the true average motion direction.

Materials and Methods

Participants

Thirty participants (13 males, age mean = 21.93 years, SD = 4.40 years) volunteered to participate in the experiment after providing informed consent. All participants had normal or corrected-to-normal vision. The study was approved by the University of Queensland Human Research Ethics Committee, and the experiment was conducted in accordance with the relevant guidelines and regulations. Participants were paid $40 for attendance. Two participants were excluded from analysis because of technical problems with the EEG recordings.

Task overview

Participants were tasked with reporting the average direction of motion across a field of moving bars (Fig. 1d; Movie 1). The field was composed of two sets of bars oriented at −45° and 45°, which each moved in different average directions. Thus, the true average motion direction across all bars was the average of the motion directions of each set of bars. We recorded EEG data while participants performed the task, using SSVEPs as a measure of attentional selectivity to deliver neurofeedback in real time. The motion coherence of both sets of moving bars varied continuously, contingent on attentional selectivity for one orientation (the trained orientation) over the other (the untrained orientation). From the participants' perspective, both sets of moving bars were equally relevant for the task. However, unbeknownst to them, higher attentional selectivity for the trained orientation resulted in greater motion coherence for both sets of bars, making the averaging task easier (Fig. 1a). Conversely, lower attentional selectivity resulted in lower motion coherence for both sets of bars, making the averaging task more difficult (Fig. 1b). Crucially, at any point in time, bars at the trained and untrained orientations moved with equal coherence; thus, there was never any physical difference in signal quality between the sets. The trained orientation was switched halfway through the experiment, such that selectivity for one set of bars was encouraged in the first half of the experiment, and selectivity for the other set of bars was encouraged for the second half of the experiment.

Movie 1.

Examples of experimental displays during neurofeedback and test trials. For each trial type, we show an example of the motion stimulus, followed by the participant's response indicating their judgment of the average motion direction, and then a behavioral feedback display indicating the true-average motion direction.

Experiment protocol

Individual trial design

On each trial, the moving bars were displayed for 3 s. SSVEPs were calculated over a 1 s sliding window, so that bar movement was completely random for the first second of each trial (Fig. 1c). Neurofeedback was implemented over the next 2 s of the trial (Fig. 1d). Movement coherence was calculated frame by frame (120 Hz refresh rate), based on the SSVEP selectivity for the trained orientation over the untrained orientation. At the end of each trial, participants used a computer mouse to rotate a dial toward the perceived average motion direction (Fig. 1d; Movie 1). To avoid biasing participants toward any particular motion direction based on the initial dial position, the dial appeared in the direction of the first mouse displacement following response cue onset. A feedback line indicating the true average motion direction appeared for 1 s immediately after participants clicked the mouse button to register a response.

Block design and counterbalancing

The experiment consisted of 12 blocks of 60 trials. On 72 of the 720 trials (10%), no neurofeedback was presented, and the motion coherence remained constant at 50% of the participant's calculated threshold. These test trials were presented pseudo-randomly, such that one test trial was presented in every group of 10 trials. During neurofeedback, motion coherence constantly shifted, contingent on the current attentional state. Any changes in SSVEP amplitude during this feedback could be attributed to these changes in the display. However, if a shift in feature-based attention was entrained by the feedback, it should emerge on trials in which there was no change in motion coherence and the displays were effectively identical for the trained and untrained orientations. The inclusion of test trials therefore allowed us to assess whether neurofeedback was effective in inducing a shift in feature-based attention. To overcome any additional noise added by the participant's preexisting biases toward one orientation over another, we used a within-subjects design, sequentially training each participant on both orientations. During the first half of the experiment, neurofeedback encouraged attention to bars of one orientation. The other orientation was trained during the second half of the experiment. Other than the changed feedback contingency, the experiment was identical in the first and second halves, and no indication was given to participants that anything had changed. The order of training of the two orientations was counterbalanced between participants, such that −45° was trained first for half of the participants and 45° was trained first for the other half. Both experimenter and participant were blind as to which orientation was trained in the first and second halves of the experiment. Each of the 16 possible directions of motion was presented on 45 trials in total. These trials were randomly interleaved throughout the experiment. Orientation and motion direction were fully counterbalanced, such that neurofeedback encouraged attention to the trained orientation rather than to any particular motion direction.

Frequency tagging design

All moving bars oscillated between black (RGB: 0, 0, 0) and white (RGB: 255, 255, 255) in a sinusoidal flicker with a 50% duty cycle to induce frequency tags. All bars of each orientation (−45°, 45°) flickered at the same frequency (13 Hz, 15 Hz). These frequencies were chosen to be beyond the range of endogenous oscillations in the theta and α ranges, which are more likely to show large variations in amplitude with visual stimulation (Demiralp et al., 2007; Rajagovindan and Ding, 2011; Renton et al., 2019a). The flicker frequencies were counterbalanced between the trained and untrained orientations, separately for neurofeedback and test trials. For each block, the trained orientation flickered at 13 Hz and the untrained orientation flickered at 15 Hz, for 50% of neurofeedback trials and 50% of test trials; on the remainder of the trials, the trained orientation flickered at 15 Hz, and the untrained orientation flickered at 13 Hz. Thus, neurofeedback encouraged attention toward the trained orientation rather than any particular flicker frequency.

Stimulus design

The moving bars subtended 0.06° × 0.19° of visual angle, and moved at a rate of 1.19° of visual angle/s across the display, within a circular aperture with a radius of 4.46° of visual angle. Bars that moved beyond the boundary of this circular area immediately reappeared at 180.00° of polar angle from the point of disappearance, maintaining the same angular trajectory. There were 400 moving bars of each orientation (−45.00°, 45.00°), and thus 800 moving bars in total. The starting position of each bar was randomly selected on each trial, with the constraint that it was within the circular display area. The drawing order of all 800 bars was also randomized trial-by-trial, such that any flickering bar was as likely to be above the others as below. This ensured that the two flicker frequencies were equally salient. Over the course of the experiment, the bars moved in 16 possible directions of motion (0.00°, 22.50°, 45.00°, 67.50°, 90.00°, 112.50°, 135.00°, 157.50°, 180.00°, 202.50°, 225.00°, 247.50°, 270.00°, 292.50°, 315.00°, 337.50°). On any given trial, the two directions of motion presented were always 67.50° apart. Thus, there were 16 possible average directions of motion (11.25°, 33.75°, 56.25°, 78.75°, 101.25°, 123.75°, 146.25°, 168.75°, 191.25°, 213.75°, 236.25°, 258.75°, 281.25°, 303.75°, 326.25°, 348.75°). A red (RGB: 255, 0, 0) fixation cross, which subtended 0.12° × 0.12° of visual angle, was presented at the center of the display, on top of all the moving bars. At the beginning of each block of 60 trials, a black (RGB: 0, 0, 0) screen announced which block would follow in white (RGB: 255, 255, 255), 40 point, Arial font (e.g., “BLOCK 1”). After an enforced 10 s break, the word “[ENTER]” appeared in 20 point font below the block number to announce that participants could proceed to the next block by a button press when ready. After each trial, participants indicated the average direction of motion across all moving bars. The response screen was black (RGB: 0, 0, 0) and contained a white (RGB: 255, 255, 255) circle with a radius of 4.46° of visual angle. This circle outlined the area within which the bars had moved during the trial.

Display computer specifications

The display was presented at a viewing distance of 57 cm on a 24-inch ASUS VG248QE monitor with a refresh rate of 120 Hz and resolution of 1920 × 1080. Stimuli were presented using the Psychophysics Toolbox (Brainard and vision, 1997) running in MATLAB R2017a (64 bit) under Windows 10 (64 bit). The experiment was run on a Dell Precision Tower 5810 desktop computer containing an Intel Xeon E7-4809 v2 CPU and NVIDIA QUADRO M4000 GPU.

EEG recording and processing

EEG recording

EEG data were sampled at 1200 Hz using a g.USBamp amplifier (g.tec Medical Engineering) from 20 active Ag/AgCl scalp electrodes arranged according to the international standard 10-20 system for electrode placement in a nylon head cap (Oostenveld and Praamstra, 2001). The electrode positions were as follows: Iz, O1, O10, O2, O9, Oz, P10, P3, P4, P7, P8, P9, PO10, PO3, PO4, PO7, PO8, PO9, POz, and Pz. This group of electrodes was chosen because SSVEPs are most strongly represented at occipitoparietal electrode sites (e.g., Renton et al., 2019b). The ground electrode was positioned at FCz, and an active Ag/AgCl ear clip electrode was used as the reference. EEG data were filtered in real time with a notch filter at 48-52 Hz and a 1-100 Hz bandpass filter. The experiment was run across three instances of MATLAB. EEG data were originally read into the first instance, then sent to a second instance for real-time processing and classification. The resulting data stream representing selectivity for the trained stimulus was then sent to the third instance of MATLAB from which the experimental task was displayed. EEG data were streamed between MATLAB instances/threads using the FieldTrip memory buffer (Oostenveld et al., 2011).

Neurofeedback

Motion coherence

Neurofeedback was implemented by manipulating the motion coherence of the moving bars. To create an average direction of motion across many moving bars, the direction of motion of each moving line was drawn from a normally distributed probability density function, with its center at the average direction of motion. The SD of this distribution of motion directions was manipulated to change the motion coherence (Waskom et al., 2018). Thus, motion was highly coherent when this distribution was narrow (Fig. 1a), and low in coherence when this distribution was broad (Fig. 1b). Neurofeedback was implemented to implicitly “reward” participants for attending to the trained orientation. Thus, as attentional selectivity for the trained orientation increased, the motion coherence of both fields of bars increased, making the task easier (Fig. 2a). The SD of the distribution of motion directions was updated every 54 ms.

Figure 2.

Real-time neurofeedback loop used to bias feature-based attention toward the trained orientation. a, The left- and right-tilted bars each flickered with different frequencies (example: blue represents 13 Hz; green represents 15 Hz), to induce SSVEPs, recorded using EEG. b, SSVEP selectivity for the trained orientation was calculated over a 1 s sliding window using the average EEG trace across the four channels that had the highest SSVEP amplitudes during a pre-experiment staircase procedure. c, The 1 s EEG epoch was submitted to an FFT, and SSVEP amplitude was taken at 13 and 15 Hz. d, The resulting SSVEP amplitudes at each frequency were compared with the population of every SSVEP amplitude for that frequency calculated for the experiment so far. e, The range of SSVEP amplitudes varies by frequency and individual; thus, the frequency assigned to the trained and untrained orientation was varied trial-by-trial, and SSVEP amplitudes were Z-transformed using the population of all SSVEP amplitudes that had been calculated for the participant at the given frequency. f, After SSVEP amplitudes were normalized, attentional selectivity was calculated as the difference between the Z scores for the trained and untrained orientations. g, The resulting measure of attentional selectivity was linearly transformed into motion coherence, such that the selectivity score mapped directly onto the SD of bar motion directions. A staircase procedure was used before the experiment to determine the motion coherence threshold at which each individual participant failed to make an accurate response on 50% of trials. Attentional selectivity of zero was aligned with this motion coherence threshold.

Real-time SSVEP selectivity

Attentional selectivity was calculated as the difference in the Z scores of the SSVEPs (13 Hz, 15 Hz) to the trained and untrained orientations. We used this difference score, rather than the absolute amplitude of the SSVEP to the trained orientation, to specifically target a relative shift in attention between the two orientations rather than the participant's arousal or sustained attention to the task. Each iteration of the real-time data analysis loop (computation speed: 54 ms; ∼19 Hz), read the last 1 s of EEG data from a memory buffer, performed an FFT (Fig. 2b,c), normalized the SSVEPs, and wrote the resulting attentional selectivity scores to a second memory buffer accessed by the task presentation thread. Before the experimental task began, the EEG data recorded during the staircase procedure (described below) were analyzed to find the four electrodes for which SSVEP amplitudes were highest at each flicker frequency. The real-time SSVEP amplitudes at the two flicker frequencies were taken as the average FFT amplitude across these four predefined best electrodes at the first harmonic (f₀) of the frequency of interest. SSVEP amplitudes were continuously added to a distribution of SSVEP amplitudes for each frequency over the course of the experiment (Fig. 2d). As flicker frequency was fully counterbalanced with bar orientation, these distributions contained a roughly even number of samples from when the SSVEP represented the trained and untrained orientation. After each sample, SSVEP amplitudes were added to frequency-specific SSVEP amplitude population distributions, and each sample was transformed to a Z score using the distribution mean (µ) and SD (σ, Fig. 2e). The population of SSVEP amplitudes was initially built up during the staircase procedure, and thus already contained 80 trials worth of samples on the first trial of the experiment. This ensured that the distribution provided a reliable estimate of the population distributions from the beginning of the neurofeedback. Once the Z scores for the SSVEPs to both flicker frequencies were calculated, the Z score for the untrained orientation was subtracted from that for the trained orientation (Fig. 2f). This difference in Z scores was linearly translated to motion coherence (Fig. 2g). When the Z score difference was zero (i.e., no selectivity for the trained or untrained motion direction), motion coherence was set to 80% of the threshold at which participants could no longer perform the task. Motion coherence was scaled around this threshold such that a Z score difference of >2.5 resulted in motion coherence with an SD of 0 (i.e., 100% motion coherence).

Staircase procedure

A staircase procedure was implemented to determine the threshold of motion coherence around which participants were unable to perform the averaging task. The staircase procedure established a baseline performance threshold, at which participants were within 67.5° of the average direction of movement on 50% of trials. Neurofeedback was implemented as a normal distribution centered at 80% of this threshold. The procedure involved four concurrent staircases, starting with the motion coherence SDs of 1.0°, 40.0°, 80.0°, and 120.0°. Trials lasted 2 s and were coded as incorrect if the participant's answer was > 67.5° from the true average. Each individual staircase was run for 20 trials, resulting in 80 staircase trials altogether. The threshold was taken as the average result across the four staircases. The mean threshold SD of dot motion directions was 41.36° (95% CI: 35.65°-48.07°, range 10°-76°).

EEG preprocessing for offline analyses

For offline analysis, noisy electrodes, identified via visual inspection, were replaced with a linear interpolation based on the nearest channels. For each trial, EEG data across all 20 channels were linearly detrended and baseline-corrected. All frequency-based analyses were performed using the average of the four EEG channels that showed the highest SSVEP amplitude for each participant and flicker frequency (13 Hz, 15 Hz). To assess which EEG channels to use for the SSVEP analyses, we computed grand-average ERPs for each participant by averaging the EEG data across every 3 s trial at each EEG channel. These ERPs were submitted to FFTs, and the four electrodes with the highest amplitudes for each frequency (13 Hz, 15 Hz) were used for all SSVEP analyses. The electrodes included in this group for the 13 Hz SSVEP, listed from most common to least common, were as follows: Oz (19 participants, -19p), O2 (17p), POz (16p), O1 (14p), PO8 (9p), PO4 (9p), Pz (7p), PO3 (6p), P3 (5p), P4 (4p), Iz (2p), O9 (1p), PO7 (1p), O10 (1p), P8 (1p). The electrodes included in this group for the 15 Hz SSVEP, listed from most common to least common, were as follows: Oz (18p), POz (17p), O2 (14p), O1 (11p), PO4 (11p), PO3 (9p), Pz (6p), PO8 (5p), O9 (3p), PO7 (3p), P4 (3p), P7 (2p), P3 (2p), O10 (2p), Iz (2p), P9 (1p), PO10 (1p), P10 (1p), P8 (1p). Trials on which the absolute amplitude at any of these channels exceeded 200 µV were excluded from further analysis. As frequency and orientation were counterbalanced and orthogonalized, this approach optimized SSVEP amplitudes equally for the trained and untrained orientations.

To assess the time course of SSVEP selectivity for the trained over the untrained orientation, we performed a time-frequency analysis. For this analysis, we generated two ERPs for each participant: one for trials in which the trained and untrained orientations flickered at 13 and 15 Hz, respectively, and one for the reverse case. These ERPs included an additional 2 s of data on either side of each trial to preserve precision at the trial beginning and end. We applied Morlet wavelet time-frequency transforms to each of these ERPs, extracting SSVEPs at both 13 and 15 Hz (Brainstorm) (Tadel et al., 2011). These transformations used mother wavelets at each target frequency with FWHM set to 0.50 s, fixing temporal resolution at 0.20 s and spectral resolution at 0.80 Hz for each frequency. The resulting SSVEPs were averaged across the two flicker frequencies to output the average time course of the neural response to the trained and untrained orientations. This approach was applied separately for neurofeedback and test trials.

Statistical tests and models

All statistical tests were performed using MATLAB (2016b). Circular statistics were computed using “Circstat: a Matlab toolbox for circular statistics” (Berens, 2009) (https://github.com/circstat/circstat-matlab). Bayes factors (BFs) were computed using the “bayesFactor” package for MATLAB (Rouder et al., 2009; Morey and Rouder, 2011; Krekelberg, 2021) (https://github.com/klabhub/bayesFactor).

Permutation tests

Time-series effects were assessed statistically using Monte Carlo permutation simulations (Chung and Romano, 2013). For the time-frequency analyses, we sought to identify periods during which SSVEPs to the trained and untrained motion directions were significantly different. For this analysis, we generated two ERPs for each participant: one for trials in which the trained and untrained orientation flickered at 13 and 15 Hz, respectively, and one for the reverse case. We applied Morlet wavelet transforms to these ERPs to assess the average time course of the neural response to the trained and untrained orientations (see EEG preprocessing for offline analyses). The difference between these two time-series was taken to reflect SSVEP selectivity for the trained orientation (SSVEP_trained – SSVEP_untrained). We implemented a Monte Carlo simulation to generate a permutation distribution of the possible values for this SSVEP selectivity score that would be expected by chance. Over 1000 iterations, we shuffled the labels for the flicker frequency of the trained and untrained orientations on individual trials, before generating the ERPs for the time-frequency analysis. This resulted in 1000 permuted time-series' of SSVEP selectivity per participant.

For the forward encoding analysis, we sought to identify periods of significant tuning to the trained, untrained, and average motion directions. For these analyses, we trained a forward encoding model based on idealized tuning curves centered at each of the possible motion directions (0°-360°). For each trial, we rotated the labels for the predicted channel responses relative to the true motion direction presented on that trial (−180° to 180°). For example, if the presented motion direction on a given trial was 180°, we would adjust the labels for this trial such that channel responses for 0° were relabeled −180° and channel responses at 360° were relabeled 180°. Tuning metrics were derived as the vector sum of channel responses aligned to the cardinal axis of the encoded motion direction (see Forward encoding model). In other words, the tuning metric represented the motion energy of channel responses in the direction of the encoded motion direction. We implemented a Monte Carlo simulation to generate a permutation distribution of possible values for this tuning metric that would be expected by chance. Over 1000 permutations, we shuffled the labels for the true motion direction presented on each trial, such that the tuning metric was calculated around different combinations of random angles. This resulted in 1000 permuted time courses of tuning to the encoded motion direction per participant.

For each of these metrics, we averaged the true and permuted time-series data across participants to identify statistically significant periods across the grand-average time-series. This averaging across participants was performed in a second permutation step, in which we generated 1,000,000 grand-average permuted time courses by averaging different combinations of the individual participants, permuted time courses on each iteration. Each permuted group average time course represented different combinations of each of the 28 participants. 1000 permuted time courses. Every participant was still included in every permuted average; we simply sampled a different one of the participant's permuted time courses on each permutation iteration. To control for multiple comparisons, the time-series data from these 1,000,000 grand averages were then all grouped within a single distribution to calculate the thresholds for significance. Using a criterion of p < 0.05 (two-tailed) on this full distribution, we designated points on the true grand-average time course as significantly different from chance if they fell below the 2.5th percentile or above the 97.5th percentile on the distribution of the permuted data.

Forward encoding model

Forward encoding models are based on the premise that EEG data should reflect a weighted combination of tuned neuronal responses. This principle can be used to estimate the strength of tuning to certain motion directions using weight matrices, which have in turn been estimated using idealized tuning curves (Brouwer and Heeger, 2009; Garcia et al., 2013; Smout et al., 2019; Tang et al., 2020). This approach allowed us to estimate the strength of tuning to the motion directions of the trained and untrained orientations as well as the grand-average motion direction over the course of each 3 s trial. To prepare the data for the forward encoding analysis, we submitted single-trial EEG data from each 3 s trial to a Morlet wavelet transform, with the FWHM set to 0.2, to maximize temporal resolution. The time-frequency transforms at the average of the second harmonic frequencies (26 Hz, 30 Hz) were used to assess how well the motion directions were encoded over time during the trial (Garcia et al., 2013). The model was evaluated using k-fold cross-validation; thus, the data were iteratively separated into train (B₁) and test (B₂) sets, such that each of the 12 experimental trial blocks served as a test set in turn (Fig. 3a).

Figure 3.

Schematic representation of forward encoding analysis. a, Single-trial EEG data from each trial, at each electrode, were submitted to Morlet wavelet transforms. The forward encoding analysis was performed iteratively over time using a 30 ms average sliding window of the time-frequency transformed data (β, illustrated by gray rectangle). b, The forward encoding analysis relied on a basis set of idealized channel outputs tuned to each of the motion directions presented. c, The forward encoding analysis relied on the premise that each block of EEG data (β, data from each electrode and trial at a specific time in the trial) should be related to the basis set of idealized channel outputs for the motion directions presented on each trial by a weight matrix (W). We therefore used linear regression to estimate the weight matrix, leaving one block of trials out. d, To decode how strongly the presented motion direction was represented on each trial, we applied the estimated weight matrix to the trial data, resulting in an estimated channel output. Critically, we iterated this process for each block of trials, such that the estimated trials were always left out of the forward encoding process used to estimate the weight matrix.

To generate an idealized set of channel outputs tuned to each motion direction, we populated a vector with the 16 possible motion directions for the trained or untrained motion direction (0.0°-337.5°, in 22.5° increments), or for the average motion directions (33.75°-371.25°, in 22.5° increments), depending on the motion direction currently encoded. For each trial, we subtracted the presented motion direction from this vector, such that the position corresponding to the presented motion direction was set to 0 (i.e., −180.0°, −157.5°, −135.0°, −112.5°, −90.0°, −67.5°, −45.0°, −22.0°, 0.0°, 22.5°, 45.0°, 67.5°, 90.00°, 112.5°, 135.0°, 157.5°). The cosine of this vector, raised to the seventh power, formed a biologically plausible set of channel outputs (C₁) representing the idealized neuronal response to each motion direction (Garcia et al., 2013). Negative values in this design matrix were set to zero (Fig. 3b). Training and testing were performed independently across time using a 20 ms average sliding window, sampled at 100 Hz. This meant that each iteration of training occurred using a single data matrix (B₁, m electrodes × n trials), which should be related to the hypothetical channel outputs (C₁, k channels × n trials) by a weight matrix (W, m electrodes × k channels) as follows: B1=WC1(1)

This weight matrix was estimated, using linear regression, as follows (Fig. 3c): Ŵ=B1C1T(C1C1T)−1(2)

And the channel responses (C₂) for the test data (B₂) were in turn based on the estimated weight matrix (Fig. 3d) as follows: Ĉ2=(ŴTŴ)−1ŴTB2(3)

Channel responses for each trial in the experiment were estimated as each experimental block was tested in turn. This process was performed 3 times for each participant: once selecting the trained motion direction on each trial, once selecting the untrained motion direction, and once for the grand-average motion direction. These responses for each trial were rearranged such that the encoded motion directions were centered in the vector of channel responses. Thus, 0°, set as the central position in the matrix of estimated channel responses, always represented the actual motion direction presented on each trial. Further, when we encoded the grand mean motion direction, we flipped the matrix of estimated channel responses such that the trained motion direction was always to the right of the average, and the untrained motion direction was always to the left. Once each trial was recentered, the average channel responses to the trained, untrained, and grand-average motion directions were calculated for each participant. These average responses were smoothed using a sliding average (80 sample window; MATLAB movmean function).

To assess how strongly the trained, untrained, and average motion directions were encoded throughout the trial, we calculated a single metric representing the degree of tuning at each time point. Channel responses were represented as polar vectors, where each vector's angle was the corrected motion direction (−180° to 180°), and each vector's amplitude was the channel response at that motion direction. We calculated the x and y components of each of these vectors. The x components are oriented in the direction of the encoded motion direction (0°). Thus, adding the x components of vectors corresponding to each motion direction resulted in a metric that represents the net response to the tuned motion direction.

Mixture distribution modeling of behavioral responses

To assess how attentional selectivity affected response bias during integrative decision-making, we first sought to identify guesses and target reversals. To this end, we fit a mixture distribution model for each participant's distribution of response errors (uncorrected, i.e., trained feature equally likely to fall on either side of the average) to describe these as a probabilistic mixture of noisy target responses (centered around true average), nontarget responses (centered 180° from true average), and guesses (uniform distribution at ∼360°; www.paulbays.com/toolbox/) (Bays et al., 2009; Schneegans and Bays, 2016). One participant was excluded from behavioral data analyses for poor performance, as their responses were best modeled by a uniform (i.e., flat) distribution from −180° to 180° relative to the correct response.

Participants' estimates of average motion direction on each trial (10) should reflect the weighted average of the two presented motion directions. To quantify these decision weights for the trained and untrained motion directions, a multiple-regression model with a term for each of the presented motion directions was fitted to each participant's target responses using ordinary least squares as follows: W=xYT(YYT)−1 where W reflects the decision weight for the trained and untrained motion direction, x reflects response error relative to the average motion direction on each trial (Δ°), and Y reflects the distance of the trained and untrained motion directions from the average motion direction on each trial (Δ°). To account for the circular nature of the data, both x and Y were expressed as complex numbers (i.e., as Cartesian coordinates) (Brouwer and Heeger, 2009; Rangelov et al., 2020).

Source code and data availability

Source code for the experimental task is as follows: https://github.com/air2310/Neurofeedback-Investigations-of-Integrative-Decision-Making. Data are available on request from corresponding author.

Results

Implicit training: debriefing questionnaire

The neurofeedback protocol embedded within the integrative decision-making task aimed to implicitly shift attention toward each of two orientations in turn. Participants were informed that the display might change over the course of the experiment depending on the EEG recording, but were not provided with any further detail. To probe whether or not participants had guessed the calculation underlying neurofeedback, we used a debriefing questionnaire after the experiment. Questions included: “Please describe any strategies you used to do the task, and whether they changed over the course of the experiment.” and “Was there anything you did that made the task easier or more difficult?” Participants typically described their strategies with phrases, such as “blurring my eyes,” “focusing on the fixation cross,” and “following the overall flow.” No participant described focusing more on bars of one orientation than another. This suggests that participants were not consciously aware that the task difficulty was dependent on their attention to one set of bars over another and that neurofeedback was indeed delivered implicitly. As the trained orientation was counterbalanced within subjects and the experimenter was blind to the training order, this suggests that any differences in the processing of the trained and untrained stimuli cannot be attributed to subject or experimenter biases and must result from neurofeedback training.

Sensory processing biases: neurofeedback training of attention

Frequency tagging

Before assessing for differences between SSVEPs to the trained and untrained orientations, we first confirmed that frequency tagging elicited reliable SSVEPs by calculating neural responses to the individual flicker frequencies. To this end, we calculated the grand mean ERP across all neurofeedback and test trials. Taking the FFT of these grand mean ERPs revealed that, overall, SSVEPs were evident at the flicker frequencies, and were similar in amplitude for neurofeedback and test trials (f₁: 13 Hz; f₂ 15 Hz; Fig. 4a). SSVEPs on both neurofeedback and test trials were centered over occipitoparietal electrode sites (Fig. 4b,c), consistent with previous frequency tagging studies involving dynamic visual displays (e.g., Renton et al., 2019b). Thus, frequency tagging was successful, and neurofeedback and test trials evoked a similar neural response overall.

Figure 4.

Grand-average frequency tagging results showing bias in feature-based attention toward the trained orientation on test trials. Bars of the trained and untrained orientation flickered at different frequencies (13 Hz, 15 Hz) to evoke SSVEPs. a, Single-trial frequency spectra generated by submitting single-trial EEG data, averaged across the unique combination of the four best electrodes for each flicker frequency, to FFTs. Blue asterisks represent SSVEPs at the fundamental flicker frequencies (f₀¹ = 13 Hz, f₀² = 15 Hz) and the second harmonic (f₁¹ = 26 Hz, f₁² = 30 Hz). b, c, EEG data were recorded from 20 occipitoparietal EEG channels. The grand-average SSVEPs for (b) test and (c) neurofeedback trials are plotted at each electrode to show the topographical distribution of SSVEPs using a cubic spline interpolation of SSVEP amplitudes. d, e, Topographical distribution of SSVEP selectivity for the trained orientation over the untrained orientation for (d) test and (e) neurofeedback trials. The topography shows a cubic spline interpolation of the difference between SSVEPs evoked by the trained and untrained orientation at each electrode. f, g, Raincloud plots of grand-average SSVEPs evoked by the trained and untrained orientations on (f) test and (g) neurofeedback trials (Allen et al., 2019). Each participant's marginal means have been subtracted from their cell means before plotting, to clearly show the direction of individual effects. (All reported statistics were performed on the untransformed data.) Each participant is represented by a unique color. BF₁₀ indicates the likelihood of the data under the alternative relative to the null hypothesis.

SSVEP selectivity during test trials

On test trials, motion coherence was not contingent on real-time SSVEP feedback. Instead, coherence remained fixed at 50% of the threshold at which participants could no longer perform the task. Any difference in attentional selectivity between the trained and untrained orientations during these trials thus provided a pure index of any trained carryover effects from the neurofeedback trials. To assess whether any such attentional selectivity for the trained feature arose on either trial type, we calculated 3 s trial ERPs for each participant, separately for the neurofeedback and test trials in which the trained orientation flickered at either 13 or 15 Hz (counterbalanced across the experiment). These ERPs were submitted to FFTs, and the amplitudes at 13 and 15 Hz were averaged to output SSVEPs to the trained and untrained orientations. To assess attentional selectivity, we calculated the difference in SSVEPs to the trained and untrained orientations, separately for neurofeedback and test trials. The topographical distribution of SSVEP selectivity (ΔµV) for the trained over the untrained orientation on test trials shows attentional selectivity for the trained orientation across a number of occipitoparietal electrode sites (Fig. 4d). To assess whether this selectivity for the trained orientation was statistically significant, mean SSVEP amplitudes (µV) evoked by the trained and untrained orientation were calculated using the mean of the four electrodes with the highest SSVEP amplitudes for each flicker frequency and participant. These mean SSVEPs were submitted to a t test, which showed that SSVEPs evoked by the trained orientation (mean = 0.69, SD = 0.42) were significantly larger than those evoked by the untrained orientation (mean = 0.65, SD = 0.38, t₍₂₇₎ = 2.28, p = 0.03, BF₁₀ = 1.84; Fig. 4f). Thus, neurofeedback was successful in shifting feature-based attention toward one orientation over the other. This attentional selectivity is based on an average across the entire trial.

We next asked at which point in the 3 s trials attentional selectivity emerged (Fig. 5a). To assess the time course of attentional selectivity for the trained over the untrained orientation, we calculated the difference between the Morlet wavelet-transformed ERPs to these two signals (Fig. 5c,e). We used a permutation test based on a Monte Carlo simulation to assess whether attentional selectivity for either the trained or untrained orientation was significant at any point during the trial. The results of this analysis revealed significant attentional selectivity for the trained over the untrained orientation on test trials, which emerged in the first second after the onset of coherent motion (Fig. 5g). This attentional selectivity for the trained orientation was significant with a criterion of p < 0.05 (two-tailed) from 163 to 889 ms after the onset of coherent motion. Thus, overall, the neurofeedback-induced enhancement of the trained orientation was most prominent in the first second of coherent motion within a trial.

Figure 5.

Time-frequency analyses of SSVEPs showing feature-based attentional selectivity for the trained orientation on test and neurofeedback trials. a, b, ERPs relative to trial onset showing the SSVEP for (a) test and (b) neurofeedback trials. The distinct waveform evident within the ERP is the combined 13 and 15 Hz frequency tag. ERPs are shown for the two frequency tagging conditions (Trained: 13 Hz, Untrained: 15 Hz; Trained: 15 Hz, Untrained: 13 Hz). c, d, Morlet wavelet transforms of the ERPs above, showing the amplitude of SSVEPs to the trained orientation over the course of (c) test and (d) neurofeedback trials. e, f, Morlet wavelet transforms of the ERPs above showing the amplitude of SSVEPs to the untrained orientation over the course of (e) test and (f) neurofeedback trials. g, h, Difference in Morlet wavelet transforms showing attentional selectivity for the trained over the untrained orientation on (g) test and (h) neurofeedback trials. Green shading represents the trial period over which selectivity for the trained orientation was significantly different from chance (p < 0.05).

SSVEP selectivity during neurofeedback training

On neurofeedback trials, motion coherence varied continuously, contingent on real-time attentional selectivity for one set of oriented bars over the other. Thus, motion coherence was higher when the trained orientation was attended and lower when the untrained orientation was attended. On these neurofeedback trials, the topographical distribution of the difference between SSVEPs to the trained and untrained orientation (ΔµV) revealed no effect at any electrode site (Fig. 4e). To assess any possible difference statistically, mean SSVEP amplitudes evoked by the trained and untrained orientations were calculated using the mean of the four electrodes with the highest SSVEP amplitudes for each flicker frequency and participant. These mean SSVEP amplitudes (µV) were submitted to a paired-samples t test, which revealed no significant difference between SSVEPs evoked by the trained (mean = 0.65, SD = 0.40) and untrained orientations during neurofeedback trials (mean = 0.66, SD = 0.41, t₍₂₇₎ = −0.99, p = 0.33, BF₁₀ = 0.31; Fig. 4g).

We next asked whether attentional selectivity emerged at any point in the 3 s trial. To do this, we performed a time-frequency analysis of the neurofeedback trial ERPs (Fig. 5b). Using the same procedure as for the test trials, we used Morlet wavelet transforms to calculate the average time course of the neural response to the trained and untrained orientations (Fig. 5d,f), and determined the difference between these two SSVEPs (Fig. 5h). To assess whether attentional selectivity for either the trained or untrained orientation was significant at any point during the trial, we used a permutation test based on a Monte Carlo simulation. The results of this analysis revealed no significant attentional selectivity for either orientation at any point during the trial (Fig. 5h). It should be noted that this does not indicate that participants never switched their attention between stimuli, but rather suggests that there was no time during which participants reliably attended more to one feature than another.

Thus, there was no overall difference in attentional selectivity for the trained and untrained orientations during neurofeedback trials. Recall, however, that SSVEPs during these neurofeedback trials were influenced by dynamic changes in motion coherence, and thus cannot unambiguously determine any stable attentional biases in favor of one orientation over the other. The test trials, in which neurofeedback was removed and thus motion coherence was stable, revealed an attentional bias. We next asked how this bias in sensory processing affected the perceptual representation of motion direction for the two stimuli, as well as the integration of the two motion signals into a perceptual average.

Integrative decision-making processes: neural representations of the trained, untrained, and average motion directions

Forward encoding model

Recall that the participants' task was to integrate the motion directions of the two overlapping stimulus fields to produce their average motion direction. The two sets of bars always moved with identical motion coherence, and each trial was followed by feedback indicating the true average motion direction. Thus, an ideal observer should weigh the two sets of bars equally in the grand average (Rahnev and Denison, 2018). However, the neurofeedback embedded within this integrated decision-making task was successful in training participants to prioritize the processing of the trained set of bars over the untrained set of bars, as evidenced by the SSVEP bias observed on the test trials. This gives us the opportunity to observe how any perceptual biases might arise, independent of any bottom-up differences in the integrated stimuli. To this end, we set out to visualize and describe the temporal dynamics of evidence accumulation for the trained, untrained, and grand-average motion directions over the course of the trial. As continuous EEG data were recorded throughout the experiment, we used a forward encoding approach in which we trained separate models for the motion directions of the two presented stimulus fields, and a further model for the average motion direction.

Forward encoding results: trained and untrained motion directions

Before assessing how the two motion directions were integrated into an average, we first assessed how the representation of the trained and untrained motion directions built up separately during the trial. Figure 6a and b show the average strength of encoding at each channel, centered around motion directions for the trained (Fig. 6a) and untrained (Fig. 6b) orientations, throughout the 2 s of coherent motion. These surface plots reveal that the motion direction of the trained orientation was strongly encoded early in the trial, and built up from the moment of coherent motion onset (Fig. 6a). This strong representation of the trained motion direction early in the trial tapered off after ∼1 s of coherent motion. By contrast, encoding to the untrained motion direction was suppressed early in the trial, but with a peak emerging toward the end of the trial (Fig. 6b).

Figure 6.

Forward encoding results showing dynamics of encoding of the trained and untrained motion directions, as well as the true average motion direction. a, b, Surface plots showing forward encoding of the component motion directions centered around the (a) trained and (b) untrained motion directions, cropped to show the 2 s of coherent motion during each trial. Cutaway panels show the average channel responses for slices from 1.00 s to 1.67 s, 1.67 s to 2.33 s, and 2.33 to 3.00 s, to illustrate how the tuned response changed over time. c, Surface plot showing forward encoding of results for the average motion direction, cropped to show the 2 s of coherent motion during each trial. Before averaging but after model training, the labels of motion directions on individual trials were flipped such that the trained motion was always at 33.75° and the untrained motion direction was always at −33.75°. Cutaway panels show the average channel responses for slices from 1.00 to 1.67 s, 1.67 to 2.33 s, and 2.33 to 3.00 s, to illustrate how the tuned response changed over time. d, Difference between tuning to trained and untrained motion directions. Tuning metrics were derived by calculating the vector sum of the forward encoding results for a particular motion direction by taking the vector components aligned with the cardinal axis of the motion direction of interest. Blue shaded area represents the 2.5th to the 97.5th percentile of a Monte Carlo distribution of the possible differences between these tuning metrics. Yellow shaded areas represent significant differences from this distribution (α = 0.05, two-tailed). e, f, Tuning metrics for the (e) trained and (f) untrained motion directions. Blue shaded areas represent the 2.5th to the 97.5th percentile of a Monte Carlo distribution of the possible values of tuning for randomly selected motion directions (chance level tuning). Yellow shaded areas represent significant differences from this distribution (α = 0.05, two-tailed). g, Tuning metric for the average motion direction. Shading is the same as for e and f.

To statistically determine how encoding of the trained and untrained motion directions differed, we computed tuning metrics to summarize the degree of tuning along the trained and untrained motion directions. We first asked when tuning to the trained and untrained motion directions differed using a permutation test based on a Monte Carlo simulation. By iteratively recalculating the mean difference in tuning to the trained and untrained motion directions with the labels for trained and untrained shuffled, we generated a permutation distribution to represent the magnitude of difference that would be expected by chance. This showed that tuning to the trained motion direction was significantly stronger than tuning to the untrained motion direction from 153 to 906 ms after the onset of coherent motion (Fig. 6d; p < 0.05). Notably, this period of significantly stronger encoding of the trained orientation overlaps the period of significant attentional selectivity for the trained orientation on test trials (Fig. 5g).

We next asked when the tuning metrics for the trained and untrained motion directions, individually, differed from what would be expected by chance. To this end, we compared the tuning metrics for the trained and untrained motion directions to permuted distributions of tuning curves created by shuffling the labels for the true motion directions presented on each trial. The permutation distribution therefore represented the range of likely tuning amplitudes that might be expected if the modeled motion direction was arbitrary. Setting α = 0.05 (two-tailed), we found that tuning to the trained motion direction was significantly stronger than chance within the first second of coherent motion, briefly from 404 to 464 ms and again from 565 to 836 ms after the onset coherent motion (Fig. 6e). By contrast, tuning to the untrained motion direction was actually weaker than might be expected by chance over the first second of the trial, from 103 to 695 ms after the onset of coherent motion. Stronger than chance tuning for this motion direction only emerged toward the end of the trial, from 1708 to 1969 ms after the onset of coherent motion (Fig. 6f). Notably, these results were generated using both the neurofeedback and test trials to maximize accuracy and reliability of the forward encoding model. However, the pattern of results described here held true for both neurofeedback trials and test trials when separated. This pattern of results suggests that, while the trained and untrained orientations were presented concurrently, they were encoded sequentially in the brain, with a strong representation of the trained motion direction prioritized because of neurofeedback.

Forward encoding results: average motion direction

We next asked how the trained and untrained motion directions were integrated into the average motion direction during the trial. Notably, this motion direction was never actually displayed during the trial but needed to be built up internally through integration of the motion directions of the two overlapping stimulus fields. Applying a forward encoding model to this motion direction allowed us to assess how neurofeedback-induced differences in the representations of the trained and untrained motion directions affected the buildup of the neural representation of the integrated average. Figure 6c shows the average strength of encoding of each motion direction channel, centered around the average motion direction, for the 2 s of coherent motion. These were flipped after model training such that the trained motion direction was always to the right of the average, and the untrained motion direction was always to the left of the average. This surface plot shows that the representation of the average motion direction built up gradually during the trial, and grew stronger until the offset of coherent motion (Fig. 6c). Notably, the peak response appears to be shifted away from the true average response toward the trained motion direction (22.5°).

In order to quantify the strength of tuning to the average motion direction over the course of the trial, we generated a summary tuning metric by calculating the vector sum of all channel responses in the direction of the encoded motion direction (i.e., channel response ×cos(θ); see Fig. 6b,c). This metric indicated the strength of tuning to the average motion direction at each point in the trial (Fig. 6g). To assess when this tuning metric differed from what might be expected by chance, we compared it with a permuted distribution of average tuning metrics generated by shuffling the labels for the true average motion directions presented on each trial. The permutation distribution therefore represented the range of likely tuning amplitudes that might be expected if the modeled motion direction was arbitrary. Setting α at 0.05 (two-tailed), we found that tuning to the true average motion direction was significantly stronger than chance from 1087 ms after the onset of coherent motion until the end of the trial. Notably, this tuning to the average motion direction occurred after the offset of significant tuning to the trained motion direction but overlapped with significant tuning to the untrained motion direction.

To investigate the shift in peak tuning away from the true average motion direction and toward the trained motion direction (Fig. 6c), we generated a summary bias metric by calculating the vector sum of all channel responses perpendicular to the direction of the average motion direction (i.e., channel response ×sin(θ)). Positive values of this metric indicate a shift in tuning toward the trained motion direction, and negative values indicate a shift in tuning toward the untrained motion direction. We used the same permutation test as that used to assess periods of significant tuning to the average motion direction to assess periods of significant shifts in tuning direction away from this average. Setting α at 0.05 (two-tailed), we found that tuning direction was shifted significantly away from the true average motion direction and toward the trained motion direction from 484-866 ms after the onset of coherent motion (corresponding with the period of significant tuning to the trained motion direction). To visualize how the amplitude and direction of tuning to the average motion direction evolved over time, we plotted the mean tuning vector for each time point after the onset of coherent motion (Movie 2). y components of these vectors were used to calculate the tuning metric, whereas the x components were used to generate the bias metric. The video shows that, during the period of significant bias, tuning was centered around the trained motion direction. Tuning then shifted toward the true average, growing in amplitude over time.

Movie 2.

Vector representation of tuning to average motion direction over the course of the trial (3 s), averaged across all participants. Average tuning vectors for each time point were derived by calculating the vector sum of forward encoding weights across motion directions. The resultant tuning vectors were averaged across participants (amplitude weighted circular mean) to find the average tuning vector across participants for each time point. These average tuning vectors are shown iteratively to illustrate how tuning evolved throughout the trial (r limits: 0.25-0.4, a.u.). Each time point is shown in a different color, with blue time points earlier in the trial and red time points later in the trial. Dashed white lines indicate the motion directions of the trained and untrained bars, as well as the average motion direction. Smaller plots to the right show the evolution of tuning to the trained, untrained, and average motion directions, and show the time corresponding with each tuning vector.

To maximize accuracy and reliability of the forward encoding model, forward encoding results were generated using both the neurofeedback and test trials. However, it should be noted that the pattern of results described here holds true for both neurofeedback and test trials when modeled separately. Thus, overall, we found that the trained orientation was encoded first, followed by the untrained orientation, before integration into a representation of the average motion direction. Further, the integrated representation of the average motion direction was shifted toward the trained motion direction. We next asked how this encoding process affected behavioral responses on each trial.

Integrative decision-making outcomes: behavioral reporting of the average motion

Having shown that neurofeedback biased the allocation of attention toward the trained motion direction, and that in turn, this motion direction was encoded earlier and more strongly than the untrained motion direction, we next asked how neurofeedback affected the behavioral outcome of the integrated decision at the end of each trial.

Response bias distribution

To visualize the distribution of response errors relative to the true average, we calculated the difference between responses and the true average motion direction on each trial, such that zero represented perfect accuracy. Mixture distribution modeling of these responses was used to separate target responses from guesses and nontarget responses (reversals; see Materials and Methods). Target responses were then transformed such that errors with positive values were closer to the trained motion direction, and errors with negative values were closer to the untrained motion direction (Fig. 7a,b). We used the circular statistics toolbox to calculate descriptive statistics for each participant's transformed distribution of response errors. On test trials, these distributions had an average circular mean of 1.28° (95% CI: −0.82° to 3.38°), circular median of 1.39° (95% CI: −1.78° to 4.56°), and mode of 0.74° (95% CI: −7.91° to 9.39°; Fig. 7a). On neurofeedback trials, these distributions had an average circular mean of −0.98° (95% CI: −2.52° to 0.56°), circular median of −0.43° (95% CI: −1.48° to 0.62), and mode of −1.67° (95% CI: −7.95° to 4.61°; Fig. 7a). Thus, participants' responses appeared to be shifted slightly toward the trained motion direction on test trials, and if anything appeared shifted toward the untrained motion direction on neurofeedback trials. We next set out to formally test how the trained and untrained motion directions were weighted in participants' responses.

Figure 7.

Behavioral results exploring the relative weighting of the trained and untrained oriented bars' motion directions in participants' responses. a, b, Polar histogram of responses indicating perceived average motion direction on (a) test trials and (b) neurofeedback trials. Responses were corrected such that 0° represents perfectly accurate responses, positive angles are closer to the trained motion direction, and negative angles are closer to the untrained motion direction. Gray dotted lines indicate the actual average motion direction of the trained and untrained orientations. Solid blue, gray, and green lines indicate the mean, median, and mode, respectively. The plotted distribution contains all responses from all participants. Descriptive statistics averaging across each individual participant's distribution of responses are provided in the text. c, d, Response bias regression weights. To assess how heavily the trained and untrained orientations were weighted in participants' reported global average motion direction, we regressed the trained and untrained motion directions against participants' responses. Using least-squares regression, we estimated the weights of the trained and untrained motion directions for (c) test trials and (d) neurofeedback trials. e, Response bias by SSVEP selectivity for test trials. Response bias (circular mean of responses) was calculated for each participant on the test trials on which SSVEP selectivity for the trained orientation was relatively high and relatively low (median split). Gray bar represents the difference in response bias for these two sets of trials for each participant (sorted). Positive response bias values (right of zero) are shifted toward the trained motion direction, whereas negative values (left of zero) are shifted toward the untrained motion direction. f, Response variance by SSVEP selectivity for neurofeedback trials. Response variance (circular SD of responses) was calculated for each participant on the neurofeedback trials on which SSVEP selectivity for the trained orientation was relatively high and relatively low (median split). For plots labeled “Subject mean subtracted,” each participant's marginal means have been subtracted from their cell means before plotting to clearly show the effect directions. (All reported statistics were performed on the untransformed data.) Each participant is represented by a unique color in each raincloud plot (Allen et al., 2019). BF₁₀ indicates the likelihood of the data under the alternative relative to the null hypothesis.

Feature decision weightings

Participants' estimates of average motion direction on each trial should reflect the weighted average of the two presented motion directions. To assess whether the trained or untrained motion direction was weighted more heavily in participants' responses, we used a regression approach to compute the least-squares estimates of the weightings of the trained and untrained orientations for each participant (see Materials and Methods). This analysis revealed that for test trials, in which participants attended significantly more strongly to the trained orientation, the trained orientation (mean = 0.69, SD = 0.31) was indeed weighted more heavily in participants' responses than the untrained orientation (mean = 0.64, SD = 0.33, t₍₂₆₎ = 2.59, p = 0.016, BF₁₀ = 3.24; Fig. 7c). By contrast, on neurofeedback trials, in which there was no overall difference in attentional selectivity for the trained and untrained orientations, weights for the trained orientation (mean = 0.56, SD = 0.33) did not differ significantly from weights for the untrained orientation (mean = 0.57, SD = 0.33; t₍₂₆₎ = −1.53, p = 0.138, BF₁₀ = 0.57; Fig. 7c).

Response bias by attentional selectivity

On test trials, participants attended more strongly to the trained feature, and in turn weighted this feature more heavily in their integrated average responses. To more directly assess this relationship, we compared response bias (°) for the trials on which SSVEP selectivity for the trained orientation was strongest, with response bias for the trials on which SSVEP selectivity was weakest. To this end, we calculated an SSVEP selectivity score (SSVEP_trained – SSVEP_untrained) for each trial by first Z scoring the single-trial SSVEPs relative to the population of the participant's single-trial SSVEPs to each flicker frequency. For each participant, we performed a median split based on these SSVEP selectivity scores (separately for neurofeedback and test trials), to obtain two equally sized samples of trials in which SSVEP selectivity for the trained orientation was at its strongest or weakest. We then calculated the circular mean of each participant's response bias scores (Δ°) for each of these groups of trials (where bias toward the trained orientation is positive). For test trials, responses were significantly more biased (°) toward the trained set of bars on trials with stronger attentional selectivity for these bars (mean = 2.95, SD = 10.65) than on trials with weaker attentional selectivity for these bars (mean = −3.21, SD = 13.25; t₍₂₆₎ = 2.21, p = 0.036, BF₁₀ = 1.62; Fig. 7e).

On neurofeedback trials, participants did not attend more to one orientation than another overall, and in turn, they weighted both sets of bars equally in their responses on these trials. Partitioning responses on these trials according to single-trial SSVEP selectivity can be used to reveal more about this relationship. At first glance, it is evident that the response bias was much more variable when SSVEP selectivity for the trained feature was weak (i.e., stronger SSVEP selectivity for untrained, mean = −1.14, SD = 8.43), than when SSVEP selectivity for the trained feature was strong (mean = −1.35, SD = 3.49). Indeed, the response bias scores failed Bartlett's test for homogeneity of variance (p < 0.001), and therefore were not submitted to a t test. To more closely evaluate this relationship, we calculated the circular SD of each participant's response bias scores for each group of trials (SSVEP selectivity for the trained orientation, relatively high vs relatively low). This showed that responses were significantly more variable when SSVEP selectivity for the trained orientation was relatively low (mean = 46.05, SD = 15.92) versus relatively high (mean = 43.15, SD = 15.98, t₍₂₆₎ = 6.15, p < 0.001, BF₁₀ = 1105; Fig. 7f). This sizable difference in response accuracy demonstrates that the neurofeedback protocol successfully made the task more difficult when participants attended to the untrained orientation, and easier when participants attended to the trained orientation.

Neurofeedback dynamics

Our neurofeedback protocol was designed to bias sensory processing toward one visual feature over another. We have shown that this protocol was successful in inducing an attentional bias which transferred to test trials, on which no neurofeedback was delivered and for which there was nothing in the display to encourage prioritized processing of one stimulus over another. To facilitate the design of future neurofeedback studies using similar protocols, we investigated the dynamics of feedback to assess how this was represented in the brain, and whether any particular features of the design might yield even stronger effects. These analyses were limited to neurofeedback trials (on which feedback was presented).

Feedback variability

Neurofeedback was designed to represent the real-time difference in attentional selectivity for the trained and untrained features. To this end, at each moment in the trial, we first calculated the SSVEPs at each of the two flicker frequencies (13 Hz, 15 Hz), which were counterbalanced trial-by-trial with orientation (trained, untrained). These SSVEPs were Z-scored relative to the population of SSVEPs at that flicker frequency to assess whether they were relatively high or low compared with what might be expected for that frequency (for an example trial, see Fig. 8a). For each time point, we calculated SSVEP selectivity for the trained feature as the difference between these two Z scores (for an example trial, see Fig. 8b). This SSVEP selectivity score was used to control the motion coherence of the moving bars, making the task easier when attentional selectivity for the trained orientation was higher (for an illustration of how motion coherence corresponded to different SSVEP selectivity scores on an example trial, see Fig. 8c).

Figure 8.

Neurofeedback dynamics. a-c, Feedback calculation over the course of a representative neurofeedback trial (Participant 1, Trial 656). a, At each time point, we began by calculating the SSVEP amplitudes at the two flicker frequencies based on a 1 s sliding window of EEG data. These SSVEP amplitudes were Z-scored relative to the population of all SSVEP amplitudes at that frequency up until that time in the experiment. The normalized SSVEP amplitudes were then labeled according to whether they were evoked by the trained or untrained orientation on this trial. b, We next calculated the difference between the normalized SSVEPs to the trained and untrained orientations to arrive at an SSVEP selectivity score that reflected how strongly participants were representing the trained orientation relative to the untrained orientation. c, These SSVEP selectivity scores were used to control motion coherence in the display. Here we selected three times during the example trial at which the SSVEP selectivity score was relatively high (Time 1), intermediate (Time 2), and low (Time 3), and created polar histograms for the trained and untrained motion directions displayed to this example participant at these time points within the trial. These have been transformed such that the average trained motion direction is at 0° and the untrained motion direction is at 67.5°. d, SD of feedback values (selectivity) across versus within neurofeedback trials. Participant marginal means have been subtracted from each participant's data to show the directions of individual effects more clearly. (All reported statistics were performed on the untransformed data.) e, SD of feedback values (selectivity) early (1-2 s) versus late (2-3 s) in neurofeedback trials. Subject means have been subtracted from each participant to more clearly show the effect. For feedback variance plots, each participant's marginal means have been subtracted from their cell means before plotting to clearly show the effect directions. In each raincloud plot, each participant is represented by a unique color (Allen et al., 2019). BF₁₀ indicates the likelihood of the data under the alternative relative to the null hypothesis.

It is worth asking whether this feedback signal represented rapid, within-trial fluctuations in the allocation of attention between the two features, or a more stable attentional state that varied across trials. To address this question, we first calculated the SD of feedback (SSVEP selectivity) values within each trial, taking the average of these values, and then calculated the average feedback values for each trial and took the SD across trials. If feedback largely reflected momentary fluctuations in attentional state, we would expect larger within-trials SDs, whereas if feedback was tracking a more slowly changing state, we would expect the SD to be larger across trials. The analysis revealed that feedback varied significantly more within trials (mean = 0.90, SD = 0.22) than across trials (mean = 0.59, SD = 0.14, t₍₂₇₎ = −13.97, p < 0.001, BF₁₀ = 9.74 × 10¹⁰; Fig. 8d), which suggests that feedback was targeting moment-by-moment differences in information processing more than gradual changes in attentional state across the experiment.

Time-frequency analysis of the test (non-neurofeedback) trials showed that attentional selectivity for the trained feature was strongest in the first half of the coherent motion period, tapering off toward the end of the trial. We next investigated whether neurofeedback was more variable earlier in the trial as well, as this would have reinforced the training contingencies more strongly during this period. To this end, we calculated the SD of feedback scores for the first half of each trial (0-1 s after the onset of coherent motion) and for the second half of each trial (1-2 s after the onset of coherent motion), and calculated the average of these SDs for each participant. Feedback was indeed more variable earlier (mean = 0.82, SD = 0.19) than later in the trial (mean = 0.78, SD = 0.18, t₍₂₇₎ = 7.2, p < 0.001, BF₁₀ = 1.56 × 10⁵; Fig. 8e). To examine feedback on an even finer timescale, we calculated the SD of feedback values across trials for each time point in the trial. For each participant, we used a standard linear regression to regress SD scores against time within the trial, and found that the slope (mean = −0.05, 95% CI: −0.08, −0.02) was significantly more negative than chance (t₍₂₇₎ = −3.02, p = 0.006, BF₁₀ = 7.63). Thus, neurofeedback was most variable soon after the onset of coherent motion, and became progressively less variable toward the end of each trial. The correspondence in timing between high variance feedback periods on neurofeedback trials and significant attentional selectivity on test trials suggests that participants may have been more strongly conditioned to attend to the trained orientation during this time.

Learning curves and switch costs

In this experiment, we implemented “bidirectional” neurofeedback; that is, participants were first trained on one orientation, then on the other, thus controlling for the trained orientation within-subjects. This design choice meant we could be certain that any neural or behavioral biases arose from neurofeedback training rather than potential preexisting orientation preferences among participants. However, this advantage was balanced against potentially attenuated training effects resulting from switching the neurofeedback contingency mid-experiment. To determine whether there were such switch costs, we first compared the training effects for the first and second halves of the experiment. We began by regressing ordinal trial number within the relevant half of the experiment (i.e., how many trials the feature had been trained for), against SSVEP selectivity on individual test (i.e., non-neurofeedback) trials, using a simple linear regression. The slope of this linear regression indicated the rate at which SSVEP selectivity for the trained orientation changed over the course of the training block. This slope value did not differ significantly from 0 in either the first half (mean = −0.02, SD = 0.19, t₍₂₇₎ = −0.52, p = 0.610, BF₁₀ = 0.23) or second half (mean = 0.03, SD = 0.20, t₍₂₇₎ = 0.73, p = 0.471, BF₁₀ = 0.26) of the experiment. Indeed, there was no significant difference in slope between experiment halves (t₍₂₇₎ = −0.79, p = 0.435, BF₁₀ = 0.27). Thus, it appears that rather than steadily increasing over the course of training, effects emerged relatively rapidly and then remained at a constant level.

To assess how switching the neurofeedback contingency affected the magnitude of these consistent training effects, we next compared SSVEP selectivity on individual test (i.e., without neurofeedback) trials across the two halves of the experiment. We found that SSVEP selectivity for the trained orientation (ΔµV) was significantly larger in the first half of the experiment (mean = 0.11, SD = 0.18) than in the second half (mean = −0.02, SD = 0.20, t₍₂₇₎ = 2.39, p = 0.024, BF₁₀ = 2.22). Thus, switching the neurofeedback contingency did indeed reduce the magnitude of the training effect. To more closely examine how SSVEP selectivity changed in response to the switched contingency, we focused on the test trials immediately preceding and following the switch. For this analysis, we split each training block into thirds, and compared the end of the first training block with the beginning and middle of the second training block. SSVEP selectivity immediately after the switch (mean = −0.13, SD = 0.29) was significantly lower than that immediately before the switch (mean = 0.13, SD = 0.29, t₍₂₇₎ = 3.78, p < 0.001, BF₁₀ = 2.22; Fig. 9). This reversal in amplitude shows that the contingency learned in the first half of the experiment carried over to the second half, even though it was no longer appropriate. However, by the middle of the second block, SSVEP selectivity had increased again (mean = 0.03, SD = 0.37, t₍₂₇₎ = 1.80, p = 0.083, BF₁₀ = 0.82), recovering to a point where it was no longer significantly different from preswitch (end of first training block) selectivity (t₍₂₇₎ = 1.15, p = 0.258, BF₁₀ = 0.37; Fig. 9). Thus, switching the feedback contingency mid-experiment did indeed dampen training effects. However, it appears that neurofeedback training was still effective in both halves of the experiment.

Figure 9.

SSVEP selectivity for the trained orientation before and after the switch in feedback contingencies. Exactly halfway through the experiment, we switched the feedback contingencies such that the orientation that had previously been designated as “trained” for feedback purposes, became “untrained,” and vice versa. This switch allowed us to balance the trained orientation within-subjects, such that any biases in attention or response could not be attributed to an individual participant's prior bias in favor of one orientation over another. To assess the impact of this switch on training effects, we compared the last two training blocks of the first contingency (Pre-switch), with the first two training blocks of the second contingency (early Post-switch), as well as the following two training blocks (late Post-switch). This granularity resulted in 24 test trials per cell. Each participant is represented by a unique color on the raincloud plot (Allen et al., 2019). Solid gray lines indicate mean values. Box plots represent quartiles. ***p < 0.001. n.s., p > 0.05.

Discussion

By using an implicit neurofeedback protocol embedded within an integrative, perceptual decision-making task, we were able to bias sensory processing in favor of one source of evidence over another, despite the sources remaining matched for reliability (Mostert et al., 2015). This unique approach revealed how integrative decision-making processes are sensitive to biases in early sensory processing stages (Andersen et al., 2011). The induced attentional bias affected the order in which different sources of evidence were encoded, as well as the strength of encoding of the two sources of evidence. As a result, sources of evidence that were prioritized in early sensory processing were weighted more heavily in behavioral reports of the integrated average.

Our neurofeedback protocol biased neural responses to visual inputs, such that the processing of one featural attribute (in this case, a specific orientation) was prioritized over a concurrent and equally salient feature (Treue and Trujillo, 1999; Saenz et al., 2002; Maunsell and Treue, 2006). This prioritization was evident in the difference between SSVEPs evoked by the different stimulus sets on test trials, even though motion coherence was matched between the stimuli and there was no benefit to prioritizing one stimulus set over the other on these trials (Müller et al., 2006; Sorger et al., 2019). Our neurofeedback protocol was unique in that it directly targeted SSVEPs: visually evoked patterns of neural activity, which scale in amplitude with attentional allocation to the evoking object (Andersen et al., 2011; Norcia et al., 2015). Neurofeedback protocols have been used to improve performance on SSVEP-based BCIs, but these protocols have targeted SSVEPs indirectly, basing feedback on endogenous α oscillations (Wan et al., 2016), or algorithm classification accuracy (Zhang et al., 2010). In the past decade, several fMRI neurofeedback studies have endeavored to train complex spatiotemporal patterns of stimulus-evoked activity, allowing researchers to directly target neural representations of visual stimuli (Shibata et al., 2011, 2016; Scharnowski et al., 2012; Nan et al., 2013; Robineau et al., 2014; deBettencourt et al., 2015; Amano et al., 2016; Cortese et al., 2016; Habes et al., 2016). Applying this approach to SSVEPs allowed us to access a near real-time readout of visual processing, and provide feedback at a much higher temporal resolution than fMRI typically allows. This high temporal resolution, together with the flexible and dynamic nature of the signal we targeted, likely accounts for the very rapid training we achieved (Wolfe and Horowitz, 2004). Many similar neurofeedback studies targeting visual processing used multiple sessions spanning several days to effect change (Zhang et al., 2010; Strehl, 2014; Wan et al., 2016; Watanabe et al., 2017). By contrast, our neurofeedback protocol was “bidirectional,” such that the same signal was both upregulated and downregulated within a single experimental session (Sorger et al., 2019). While this approach allowed us to attribute the observed training effects definitively to the neurofeedback protocol, we found that it diminished the overall magnitude of the training effects across the experiment. In future studies, it will be important to balance the advantages of such controls over experimental design against the potential for larger, and perhaps more enduring, training effects. In the present study, while the training effect was somewhat diminished by our implementation of “bidirectional” control, it was nevertheless sufficient to bias integrative decision-making.

We investigated how the neurofeedback-induced bias in sensory processing of two sources of evidence affected integrative decision-making processes. Forward encoding analyses showed that neurofeedback training resulted in changes in both the sequential order and strength of encoding of the concurrently presented motion stimuli (Eldar et al., 2018; Tang et al., 2018, 2020). Interestingly, such “serial” encoding of display elements in integrative decision-making is characteristic of more complex decisions involving the integration of many distinct sources of evidence (Eldar et al., 2018). Serial processing allows more resources to be dedicated to the processing of each stimulus, allowing the trained stimulus to be encoded not only earlier in the trial, but also more strongly than would be possible under parallel processing (Treisman, 1969; Bergen and Julesz, 1983). It is notable that these differences in the encoding of trained and untrained stimuli were evident in both neurofeedback and test trials. However, the difference in SSVEP amplitude between these stimuli was only apparent on test trials. By contrast, the forward encoding results revealed that there were critical differences in the way the two sets of stimuli were encoded even during neurofeedback (Ho et al., 2012; Scolari et al., 2012). Moreover, the fact that these differences in the representations of the two stimuli were evident in the absence of a clear SSVEP difference suggests that the encoding results are not purely the result of a shift in sensory processing first to one stimulus, then to the other (Müller et al., 2006; Ales and Norcia, 2009). Rather, the forward encoding results more likely represent differences in sensory decision processes, with the representation of the trained motion direction prioritized over that of the untrained motion direction (Mostert et al., 2015). It should be noted that neurofeedback targeted the relative neural response to bars of the two different orientations. These orientations were entirely decoupled from motion direction. Thus, the paradigm never directly rewarded the representation of any particular motion direction, but rather encouraged sensory processing of a particular source of evidence. In time, evidence from this source was encoded earlier and more strongly than evidence from the untrained source.

These differences in the order and strength of encoding of the trained and untrained motion direction were reflected in both the neural representation and behavioral reports of the integrated average. Forward encoding results showed that this neural representation built up gradually during the trial, consistent with evidence accumulation through sequential sampling (Usher and McClelland, 2001; Ratcliff and McKoon, 2008; Kelly and O'Connell, 2015; van Vugt et al., 2019). This encoded representation was shifted away from the true average and toward the trained motion direction, and this shift was also reflected in behavior. Indeed, evidence from the stimulus prioritized in early sensory processing was weighted more strongly in participants' behavioral reports of the integrated average. Further, this response bias was strongest when the attentional bias was strongest. This pattern of results is similar to that found for sources of evidence that differ in reliability. In such studies, shifts in mean responses toward more reliable sources of evidence have been attributed to more precise neural representations, leading to stronger decision-weightings for this information (Landy et al., 1995; Zylberberg et al., 2016; Rahnev and Denison, 2018). However, while it may be optimal for integrative decision-making processes to discount genuinely unreliable sources of evidence, our findings suggest that this discounting still occurs when differences in the precision of representation of two sources of evidence originate from internal biases in early sensory processing, rather than from true differences in reliability (Trommershäuser, 2009; Ma, 2010).

Integrative perceptual decision-making is an essential cognitive function, allowing us to navigate complex and dynamic visual environments. The mechanics of these decision processes and their interaction with low-level visual processing are therefore worthy of further investigation. For example, in studies of single-stimulus decision-making, perceptual biases toward rewarded response alternatives have been shown to be strongest when responses are required within the first 500 ms after stimulus onset, tapering off as stimulus presentation times approach 2 s (Gao et al., 2011). This reduction in bias with additional sampling has been explained in terms of dynamic integration of stimulus information over time, and can be well modeled by a leaky accumulator. In leaky accumulator models, evidence accumulation is subject to leakage over time, allowing for the correction of any initial biases as more evidence is accumulated (Usher and McClelland, 2001; Tsetsos et al., 2012). It would be valuable to know whether a similar model might explain evidence integration across different sources of evidence, especially given our forward encoding results, which showed that the encoded representation of the average motion direction was initially strongly biased toward the trained motion direction. Further, here we operationalized integrative decision-making as the perceptual averaging of a single stimulus feature (motion direction) across different sources of evidence. It remains unclear whether these effects would persist when integration occurs across different stimulus features, which could be represented by separate populations of neurons (Emmanouil and Treisman, 2008). Consider crossing a busy road, for example. This is a complex decision-making task that relies on the integration of multiple distinct visual features, such as the color of the traffic lights as well as the speed and trajectory of nearby vehicles. Errors during such “complex” integrative perceptual decisions could prove fatal, and it is thus important we seek a better understanding of these processes.

There are many inherent biases in how sensory processing resources are allocated across visual inputs, with differences occurring across levels of statistical regularity (Appelle, 1972), previous reward associations (Anderson, 2016; Failing and Theeuwes, 2018), and current top-down goals (Yantis, 1993; Theeuwes et al., 2010). Our findings show that such biases in sensory processing can lead to differences in the precision and timing of the representation of evidence from different sources, leading to biased integrative perceptual decision outcomes. The ability to rapidly and implicitly induce such biased sensory processing within one experimental session has potentially wide-ranging applications in both basic and translational neuroscience. Neurofeedback protocols of this nature could be applied to address unanswered questions about how visual selective attention interacts with other cognitive functions (e.g., expectation) (Bang and Rahnev, 2017), and could be harnessed for cognitive training to enhance performance in neurologic patients (Robineau et al., 2014; Sitaram et al., 2017).

Footnotes

The authors declare no competing financial interests.
J.B.M. and A.I.R were supported by the ARC Center of Excellence for Integrative Brain Function ARC Center Grant CE140100007.
Correspondence should be addressed to Angela I. Renton at angie.renton23{at}gmail.com

SfN exclusive license.

References

↵
2. Ales JM,
3. Norcia AM
(2009) Assessing direction-specific adaptation using the steady-state visual evoked potential: results from EEG source imaging. J Vis 9:8. doi:10.1167/9.7.8 pmid:19761323
OpenUrl Abstract/FREE Full Text
↵
2. Allen M,
3. Poggiali D,
4. Whitaker K,
5. Marshall TR,
6. Kievit RA
(2019) Raincloud plots: a multi-platform tool for robust data visualization. Wellcome Open Res 4:63. doi:10.12688/wellcomeopenres.15191.1 pmid:31069261
OpenUrl CrossRef PubMed
↵
2. Amano K,
3. Shibata K,
4. Kawato M,
5. Sasaki Y,
6. Watanabe T
(2016) Learning to associate orientation with color in early visual areas by associative decoded fMRI neurofeedback. Curr Biol 26:1861–1866. doi:10.1016/j.cub.2016.05.014 pmid:27374335
OpenUrl CrossRef PubMed
↵
2. Andersen SK,
3. Fuchs S,
4. Müller MM
(2011) Effects of feature-selective and spatial attention at different stages of visual processing. J Cogn Neurosci 23:238–246. doi:10.1162/jocn.2009.21328 pmid:19702461
OpenUrl CrossRef PubMed
↵
2. Anderson BA
(2016) The attention habit: how reward learning shapes attentional selection: the attention habit. Ann NY Acad Sci 1369:24–39. doi:10.1111/nyas.12957 pmid:26595376
OpenUrl CrossRef PubMed
↵
2. Appelle S
(1972) Perception and discrimination as a function of stimulus orientation: the 'oblique effect' in man and animals. Psychol Bull 78:266–278. doi:10.1037/h0033117 pmid:4562947
OpenUrl CrossRef PubMed
↵
2. Bang JW,
3. Rahnev D
(2017) Stimulus expectation alters decision criterion but not sensory signal in perceptual decision making. Sci Rep 7:17072. doi:10.1038/s41598-017-16885-2 pmid:29213117
OpenUrl CrossRef PubMed
↵
2. Bays PM,
3. Catalao RF,
4. Husain M
(2009) The precision of visual working memory is set by allocation of a shared resource. J Vis 9:7.1. doi:10.1167/9.10.7 pmid:19810788
OpenUrl CrossRef PubMed
↵
2. Berens P
(2009) CircStat: a MATLAB toolbox for circular statistics. J Stat Softw 31:1–21.
OpenUrl CrossRef PubMed
↵
2. Bergen JR,
3. Julesz B
(1983) Parallel versus serial processing in rapid pattern discrimination. Nature 303:696–698. doi:10.1038/303696a0 pmid:6855915
OpenUrl CrossRef PubMed
↵
2. Brainard DH,
3. Vision S
(1997) The psychophysics toolbox. Spatial vision 10:433–436.
OpenUrl CrossRef PubMed
↵
2. Brouwer GJ,
3. Heeger DJ
(2009) Decoding and reconstructing color from responses in human visual cortex. J Neurosci 29:13992–14003. doi:10.1523/JNEUROSCI.3577-09.2009 pmid:19890009
OpenUrl Abstract/FREE Full Text
↵
2. Brunton BW,
3. Botvinick MM,
4. Brody CD
(2013) Rats and humans can optimally accumulate evidence for decision-making. Science 340:95–98. doi:10.1126/science.1233912 pmid:23559254
OpenUrl Abstract/FREE Full Text
↵
2. Chung E,
3. Romano JP
(2013) Exact and asymptotically robust permutation tests. Ann Stat 41:484–507.
OpenUrl
↵
2. Cortese A,
3. Amano K,
4. Koizumi A,
5. Kawato M,
6. Lau H
(2016) Multivoxel neurofeedback selectively modulates confidence without changing perceptual performance. Nat Commun 7:13669. doi:10.1038/ncomms13669 pmid:27976739
OpenUrl CrossRef PubMed
↵
2. deBettencourt MT,
3. Cohen JD,
4. Lee RF,
5. Norman KA,
6. Turk-Browne NB
(2015) Closed-loop training of attention with real-time brain imaging. Nat Neurosci 18:470–475. doi:10.1038/nn.3940 pmid:25664913
OpenUrl CrossRef PubMed
↵
2. Demiralp T,
3. Bayraktaroglu Z,
4. Lenz D,
5. Junge S,
6. Busch NA,
7. Maess B,
8. Ergen M,
9. Herrmann CS
(2007) Gamma amplitudes are coupled to theta phase in human EEG during visual perception. Int J Psychophysiol 64:24–30. doi:10.1016/j.ijpsycho.2006.07.005 pmid:16956685
OpenUrl CrossRef PubMed
↵
2. Eldar E,
3. Bae GJ,
4. Kurth-Nelson Z,
5. Dayan P,
6. Dolan RJ
(2018) Magnetoencephalography decoding reveals structural differences within integrative decision processes. Nat Hum Behav 2:670–681. doi:10.1038/s41562-018-0423-3 pmid:31346283
OpenUrl CrossRef PubMed
↵
2. Emmanouil TA,
3. Treisman A
(2008) Dividing attention across feature dimensions in statistical processing of perceptual groups. Percept Psychophys 70:946–954. doi:10.3758/pp.70.6.946 pmid:18717382
OpenUrl CrossRef PubMed
↵
2. Failing M,
3. Theeuwes J
(2018) Selection history: how reward modulates selectivity of visual attention. Psychon Bull Rev 25:514–538. doi:10.3758/s13423-017-1380-y pmid:28986770
OpenUrl CrossRef PubMed
↵
2. Fetsch CR,
3. DeAngelis GC,
4. Angelaki DE
(2013) Bridging the gap between theories of sensory cue integration and the physiology of multisensory neurons. Nat Rev Neurosci 14:429–442. doi:10.1038/nrn3503 pmid:23686172
OpenUrl CrossRef PubMed
↵
2. Gao J,
3. Tortell R,
4. McClelland JL
(2011) Dynamic integration of reward and stimulus information in perceptual decision-making. PLoS One 6:e16749. doi:10.1371/journal.pone.0016749 pmid:21390225
OpenUrl CrossRef PubMed
↵
2. Garcia JO,
3. Srinivasan R,
4. Serences JT
(2013) Near-real-time feature-selective modulations in human cortex. Curr Biol 23:515–522. doi:10.1016/j.cub.2013.02.013 pmid:23477721
OpenUrl CrossRef PubMed
↵
2. Gold JI,
3. Shadlen MN
(2007) The neural basis of decision making. Annu Rev Neurosci 30:535–574. doi:10.1146/annurev.neuro.29.051605.113038 pmid:17600525
OpenUrl CrossRef PubMed
↵
2. Habes I,
3. Rushton S,
4. Johnston SJ,
5. Sokunbi MO,
6. Barawi K,
7. Brosnan M,
8. Daly T,
9. Ihssen N,
10. Linden DEJ
(2016) fMRI neurofeedback of higher visual areas and perceptual biases. Neuropsychologia 85:208–215. doi:10.1016/j.neuropsychologia.2016.03.031 pmid:27020139
OpenUrl CrossRef PubMed
↵
2. Ho T,
3. Brown S,
4. van Maanen L,
5. Forstmann BU,
6. Wagenmakers EJ,
7. Serences JT
(2012) The optimality of sensory processing during the speed–accuracy tradeoff. J Neurosci 32:7992–8003. doi:10.1523/JNEUROSCI.0340-12.2012 pmid:22674274
OpenUrl Abstract/FREE Full Text
↵
2. John-Saaltink ES,
3. Kok P,
4. Lau HC,
5. de Lange FP
(2016) Serial dependence in perceptual decisions is reflected in activity patterns in primary visual cortex. J Neurosci 36:6186–6192. doi:10.1523/JNEUROSCI.4390-15.2016 pmid:27277797
OpenUrl Abstract/FREE Full Text
↵
2. Kelly SP,
3. O'Connell RG
(2015) The neural processes underlying perceptual decision making in humans: recent progress and future directions. J Physiol Paris 109:27–37. doi:10.1016/j.jphysparis.2014.08.003 pmid:25204272
OpenUrl CrossRef PubMed
↵
2. Krekelberg B
(2021) klabhub/bayesFactor: version 2.1.0. dding log numerics and linear model analyses. Zenodo. Available at https://zenodo.org/record/4849568.
↵
2. Landy MS,
3. Maloney LT,
4. Johnston EB,
5. Young M
(1995) Measurement and modeling of depth cue combination: in defense of weak fusion. Vision Res 35:389–412. doi:10.1016/0042-6989(94)00176-m pmid:7892735
OpenUrl CrossRef PubMed
↵
2. Ma WJ
(2010) Signal detection theory, uncertainty, and Poisson-like population codes. Vision Res 50:2308–2319. doi:10.1016/j.visres.2010.08.035 pmid:20828581
OpenUrl CrossRef PubMed
↵
2. Maunsell JH,
3. Treue S
(2006) Feature-based attention in visual cortex. Trends Neurosci 29:317–322. doi:10.1016/j.tins.2006.04.001 pmid:16697058
OpenUrl CrossRef PubMed
↵
2. Morey RD,
3. Rouder JN
(2011) Bayes factor approaches for testing interval null hypotheses. Psychol Methods 16:406–419. doi:10.1037/a0024377 pmid:21787084
OpenUrl CrossRef PubMed
↵
2. Mostert P,
3. Kok P,
4. de Lange FP
(2015) Dissociating sensory from decision processes in human perceptual decision making. Sci Rep 5:18253. doi:10.1038/srep18253 pmid:26666393
OpenUrl CrossRef PubMed
↵
2. Moxon KA,
3. Foffani G
(2015) Brain-machine interfaces beyond neuroprosthetics. Neuron 86:55–67. doi:10.1016/j.neuron.2015.03.036 pmid:25856486
OpenUrl CrossRef PubMed
↵
2. Müller MM,
3. Andersen S,
4. Trujillo NJ,
5. Valdés-Sosa P,
6. Malinowski P,
7. Hillyard SA
(2006) Feature-selective attention enhances color signals in early visual areas of the human brain. Proc Natl Acad Sci USA 103:14250–14254. doi:10.1073/pnas.0606668103 pmid:16956975
OpenUrl Abstract/FREE Full Text
↵
2. Nan W,
3. Wan F,
4. Lou CI,
5. Vai MI,
6. Rosa A
(2013) Peripheral visual performance enhancement by neurofeedback training. Appl Psychophysiol Biofeedback 38:285–291. doi:10.1007/s10484-013-9233-6 pmid:24101183
OpenUrl CrossRef PubMed
↵
2. Norcia AM,
3. Appelbaum LG,
4. Ales JM,
5. Cottereau BR,
6. Rossion B
(2015) The steady-state visual evoked potential in vision research: a review. J Vis 15:4. doi:10.1167/15.6.4 pmid:26024451
OpenUrl Abstract/FREE Full Text
↵
2. Oostenveld R,
3. Praamstra P
(2001) The five percent electrode system for high-resolution EEG and ERP measurements. Clin Neurophysiol 112:713–719. doi:10.1016/s1388-2457(00)00527-7 pmid:11275545
OpenUrl CrossRef PubMed
↵
2. Oostenveld R,
3. Fries P,
4. Maris E,
5. Schoffelen JM
(2011) FieldTrip: Open Source software for advanced analysis of MEG, EEG, and invasive electrophysiological data. Comput Intell Neurosci 2011:156869. doi:10.1155/2011/156869 pmid:21253357
OpenUrl CrossRef PubMed
↵
2. Painter DR,
3. Dux PE,
4. Travis SL,
5. Mattingley JB
(2014) Neural responses to target features outside a search array are enhanced during conjunction but not unique-feature search. J Neurosci 34:3390–3401. doi:10.1523/JNEUROSCI.3630-13.2014 pmid:24573295
OpenUrl Abstract/FREE Full Text
↵
2. Purcell BA,
3. Heitz RP,
4. Cohen JY,
5. Schall JD,
6. Logan GD,
7. Palmeri TJ
(2010) Neurally constrained modeling of perceptual decision making. Psychol Rev 117:1113–1143. doi:10.1037/a0020311 pmid:20822291
OpenUrl CrossRef PubMed
↵
2. Rahnev D,
3. Denison RN
(2018) Suboptimality in perceptual decision making. Behav Brain Sci 41:1–107. doi:10.1017/S0140525X18000936 pmid:29485020
OpenUrl CrossRef PubMed
↵
2. Rajagovindan R,
3. Ding M
(2011) From prestimulus alpha oscillation to visual-evoked response: an inverted-U function and its attentional modulation. J Cogn Neurosci 23:1379–1394. doi:10.1162/jocn.2010.21478 pmid:20459310
OpenUrl CrossRef PubMed
↵
2. Rangelov D,
3. West R,
4. Mattingley JB
(2020) Stimulus reliability automatically biases temporal integration of discrete perceptual targets. bioRxiv 841353.
↵
2. Raposo D,
3. Sheppard JP,
4. Schrater PR,
5. Churchland AK
(2012) Multisensory decision-making in rats and humans. J Neurosci 32:3726–3735. doi:10.1523/JNEUROSCI.4998-11.2012 pmid:22423093
OpenUrl Abstract/FREE Full Text
↵
2. Ratcliff R,
3. McKoon G
(2008) The diffusion decision model: theory and data for two-choice decision tasks. Neural Comput 20:873–922. doi:10.1162/neco.2008.12-06-420 pmid:18085991
OpenUrl CrossRef PubMed
↵
2. Ratcliff R,
3. Rouder JN
(1998) Modeling response times for two-choice decisions. Psychol Sci 9:347–356. doi:10.1111/1467-9280.00067
OpenUrl CrossRef PubMed
↵
2. Renton AI,
3. Mattingley JB,
4. Painter DR
(2019a) Optimising non-invasive brain-computer interface systems for free communication between naïve human participants. Sci Rep 9:18705. doi:10.1038/s41598-019-55166-y pmid:31822715
OpenUrl CrossRef PubMed
↵
2. Renton AI,
3. Painter DR,
4. Mattingley JB
(2019b) Differential deployment of visual attention during interactive approach and avoidance behavior. Cereb Cortex 29:2366–2383. doi:10.1093/cercor/bhy105 pmid:29750259
OpenUrl CrossRef PubMed
↵
2. Robineau F,
3. Rieger SW,
4. Mermoud C,
5. Pichon S,
6. Koush Y,
7. Van De Ville D,
8. Vuilleumier P,
9. Scharnowski F
(2014) Self-regulation of inter-hemispheric visual cortex balance through real-time fMRI neurofeedback training. Neuroimage 100:1–14. doi:10.1016/j.neuroimage.2014.05.072 pmid:24904993
OpenUrl CrossRef PubMed
↵
2. Rouder JN,
3. Speckman PL,
4. Sun D,
5. Morey RD,
6. Iverson G
(2009) Bayesian t tests for accepting and rejecting the null hypothesis. Psychon Bull Rev 16:225–237. doi:10.3758/PBR.16.2.225 pmid:19293088
OpenUrl CrossRef PubMed
↵
2. Saenz M,
3. Buracas GT,
4. Boynton GM
(2002) Global effects of feature-based attention in human visual cortex. Nat Neurosci 5:631–632. doi:10.1038/nn876 pmid:12068304
OpenUrl CrossRef PubMed
↵
2. Salzman CD,
3. Britten KH,
4. Newsome WT
(1990) Cortical microstimulation influences perceptual judgments of motion direction. Nature 346:174–177. doi:10.1038/346174a0 pmid:2366872
OpenUrl CrossRef PubMed
↵
2. Scharnowski F,
3. Hutton C,
4. Josephs O,
5. Weiskopf N,
6. Rees G
(2012) Improving visual perception through neurofeedback. J Neurosci 32:17830–17841. doi:10.1523/JNEUROSCI.6334-11.2012 pmid:23223302
OpenUrl Abstract/FREE Full Text
↵
2. Schneegans S,
3. Bays PM
(2016) No fixed item limit in visuospatial working memory. Cortex 83:181–193. doi:10.1016/j.cortex.2016.07.021 pmid:27565636
OpenUrl CrossRef PubMed
↵
2. Scolari M,
3. Byers A,
4. Serences JT
(2012) Optimal deployment of attentional gain during fine discriminations. J Neurosci 32:7723–7733. doi:10.1523/JNEUROSCI.5558-11.2012 pmid:22649250
OpenUrl Abstract/FREE Full Text
↵
2. Shibata K,
3. Watanabe T,
4. Kawato M,
5. Sasaki Y
(2016) Differential activation patterns in the same brain region led to opposite emotional states. PLoS Biol 14:e1002546. doi:10.1371/journal.pbio.1002546 pmid:27608359
OpenUrl CrossRef PubMed
↵
2. Shibata K,
3. Watanabe T,
4. Sasaki Y,
5. Kawato M
(2011) Perceptual learning incepted by decoded fMRI neurofeedback without stimulus presentation. Science 334:1413–1415. doi:10.1126/science.1212003 pmid:22158821
OpenUrl Abstract/FREE Full Text
↵
2. Sitaram R,
3. Ros T,
4. Stoeckel L,
5. Haller S,
6. Scharnowski F,
7. Lewis-Peacock J,
8. Weiskopf N,
9. Blefari ML,
10. Rana M,
11. Oblak E,
12. Birbaumer N,
13. Sulzer J
(2017) Closed-loop brain training: the science of neurofeedback. Nat Rev Neurosci 18:86–100. doi:10.1038/nrn.2016.164 pmid:28003656
OpenUrl CrossRef PubMed
↵
2. Smout CA,
3. Tang MF,
4. Garrido MI,
5. Mattingley JB
(2019) Attention promotes the neural encoding of prediction errors. PLoS Biol 17:e2006812. doi:10.1371/journal.pbio.2006812 pmid:30811381
OpenUrl CrossRef PubMed
↵
2. Sorger B,
3. Scharnowski F,
4. Linden DE,
5. Hampson M,
6. Young KD
(2019) Control freaks: towards optimal selection of control conditions for fMRI neurofeedback studies. Neuroimage 186:256–265. doi:10.1016/j.neuroimage.2018.11.004 pmid:30423429
OpenUrl CrossRef PubMed
↵
2. Spitzer B,
3. Waschke L,
4. Summerfield C
(2017) Selective overweighting of larger magnitudes during noisy numerical comparison. Nat Hum Behav 1:145–148. doi:10.1038/s41562-017-0145 pmid:32340412
OpenUrl CrossRef PubMed
↵
2. Strehl U
(2014) What learning theories can teach us in designing neurofeedback treatments. Front Hum Neurosci 8:894. doi:10.3389/fnhum.2014.00894 pmid:25414659
OpenUrl CrossRef PubMed
↵
2. Summerfield C,
3. De Lange FP
(2014) Expectation in perceptual decision making: neural and computational mechanisms. Nat Rev Neurosci 15:745–756. doi:10.1038/nrn3838 pmid:25315388
OpenUrl CrossRef PubMed
↵
2. Summerfield C,
3. Tsetsos K
(2015) Do humans make good decisions? Trends Cogn Sci 19:27–34. doi:10.1016/j.tics.2014.11.005
OpenUrl CrossRef PubMed
↵
2. Tadel F,
3. Baillet S,
4. Mosher JC,
5. Pantazis D,
6. Leahy RM
(2011) Brainstorm: a user-friendly application for MEG/EEG analysis. Comput Intell Neurosci 2011:879716. doi:10.1155/2011/879716 pmid:21584256
OpenUrl CrossRef PubMed
↵
2. Tang MF,
3. Ford L,
4. Arabzadeh E,
5. Enns JT,
6. Visser TA,
7. Mattingley JB
(2020) Neural dynamics of the attentional blink revealed by encoding orientation selectivity during rapid visual presentation. Nat Commun 11:434. doi:10.1038/s41467-019-14107-z pmid:31974370
OpenUrl CrossRef PubMed
↵
2. Tang MF,
3. Smout CA,
4. Arabzadeh E,
5. Mattingley JB
(2018) Prediction error and repetition suppression have distinct effects on neural representations of visual information. eLife 7:e33123. doi:10.7554/eLife.33123
OpenUrl CrossRef PubMed
↵
2. Theeuwes J,
3. Olivers CNL,
4. Belopolsky A
(2010) Stimulus-driven capture and contingent capture. Wiley Interdiscip Rev Cogn Sci 1:872–881. doi:10.1002/wcs.83 pmid:26271783
OpenUrl CrossRef PubMed
↵
2. Treisman AM
(1969) Strategies and models of selective attention. Psychol Rev 76:282–299. doi:10.1037/h0027242 pmid:4893203
OpenUrl CrossRef PubMed
↵
2. Treue S,
3. Trujillo JC
(1999) Feature-based attention influences motion processing gain in macaque visual cortex. Nature 399:575–579. doi:10.1038/21176 pmid:10376597
OpenUrl CrossRef PubMed
↵
2. Trommershäuser J
(2009) Biases and optimality of sensory-motor and cognitive decisions. Prog Brain Res 174:267–278. doi:10.1016/S0079-6123(09)01321-1 pmid:19477345
OpenUrl CrossRef PubMed
↵
2. Tsetsos K,
3. Gao J,
4. McClelland JL,
5. Usher M
(2012) Using time-varying evidence to test models of decision dynamics: bounded diffusion vs. the leaky competing accumulator model. Front Neurosci 6:79. doi:10.3389/fnins.2012.00079 pmid:22701399
OpenUrl CrossRef PubMed
↵
2. Usher M,
3. McClelland JL
(2001) The time course of perceptual choice: the leaky, competing accumulator model. Psychol Rev 108:550–592. doi:10.1037/0033-295x.108.3.550 pmid:11488378
OpenUrl CrossRef PubMed
↵
2. van Vugt MK,
3. Beulen MA,
4. Taatgen NA
(2019) Relation between centro-parietal positivity and diffusion model parameters in both perceptual and memory-based decision making. Brain Res 1715:1–12. doi:10.1016/j.brainres.2019.03.008 pmid:30876858
OpenUrl CrossRef PubMed
↵
2. Wan F,
3. da Cruz JN,
4. Nan W,
5. Wong CM,
6. Vai MI,
7. Rosa A
(2016) Alpha neurofeedback training improves SSVEP-based BCI performance. J Neural Eng 13:036019. doi:10.1088/1741-2560/13/3/036019
OpenUrl CrossRef
↵
2. Waskom ML,
3. Asfour J,
4. Kiani R
(2018) Perceptual insensitivity to higher-order statistical moments of coherent random dot motion. J Vis 18:9. doi:10.1167/18.6.9 pmid:30029220
OpenUrl CrossRef PubMed
↵
2. Watanabe T,
3. Sasaki Y,
4. Shibata K,
5. Kawato M
(2017) Advances in fMRI real-time neurofeedback. Trends Cogn Sci 21:997–1010. doi:10.1016/j.tics.2017.09.010 pmid:29031663
OpenUrl CrossRef PubMed
↵
2. Wolfe JM,
3. Horowitz TS
(2004) What attributes guide the deployment of visual attention and how do they do it? Nat Rev Neurosci 5:495–501. doi:10.1038/nrn1411
OpenUrl CrossRef PubMed
↵
2. Yantis S
(1993) Stimulus-driven attentional capture. Curr Dir Psychol Sci 2:156–161. doi:10.1111/1467-8721.ep10768973
OpenUrl CrossRef
↵
2. Zhang D,
3. Maye A,
4. Gao X,
5. Hong B,
6. Engel AK,
7. Gao S
(2010) An independent brain–computer interface using covert non-spatial visual selective attention. J Neural Eng 7:16010. doi:10.1088/1741-2560/7/1/016010 pmid:20083864
OpenUrl CrossRef PubMed
↵
2. Zylberberg A,
3. Fetsch CR,
4. Shadlen MN
(2016) The influence of evidence volatility on choice, reaction time and confidence in a perceptual decision. eLife 5:e17688. doi:10.7554/eLife.17688
OpenUrl CrossRef PubMed

In this issue

View Full Page PDF

Citation Tools

Respond to this article

Request Permissions

Keywords

Cited By...

Research Articles

Show more Research Articles

Behavioral/Cognitive

Show more Behavioral/Cognitive

[1] ↵

Ales JM,
Norcia AM
(2009) Assessing direction-specific adaptation using the steady-state visual evoked potential: results from EEG source imaging. J Vis 9:8. doi:10.1167/9.7.8 pmid:19761323
OpenUrl Abstract/FREE Full Text

[3] Ales JM,

[4] Norcia AM

[5] ↵

Allen M,
Poggiali D,
Whitaker K,
Marshall TR,
Kievit RA
(2019) Raincloud plots: a multi-platform tool for robust data visualization. Wellcome Open Res 4:63. doi:10.12688/wellcomeopenres.15191.1 pmid:31069261
OpenUrl CrossRef PubMed

[7] Allen M,

[8] Poggiali D,

[9] Whitaker K,

[10] Marshall TR,

[11] Kievit RA

[12] ↵

Amano K,
Shibata K,
Kawato M,
Sasaki Y,
Watanabe T
(2016) Learning to associate orientation with color in early visual areas by associative decoded fMRI neurofeedback. Curr Biol 26:1861–1866. doi:10.1016/j.cub.2016.05.014 pmid:27374335
OpenUrl CrossRef PubMed

[14] Amano K,

[15] Shibata K,

[16] Kawato M,

[17] Sasaki Y,

[18] Watanabe T

[19] ↵

Andersen SK,
Fuchs S,
Müller MM
(2011) Effects of feature-selective and spatial attention at different stages of visual processing. J Cogn Neurosci 23:238–246. doi:10.1162/jocn.2009.21328 pmid:19702461
OpenUrl CrossRef PubMed

[21] Andersen SK,

[22] Fuchs S,

[23] Müller MM

[24] ↵

Anderson BA
(2016) The attention habit: how reward learning shapes attentional selection: the attention habit. Ann NY Acad Sci 1369:24–39. doi:10.1111/nyas.12957 pmid:26595376
OpenUrl CrossRef PubMed

[26] Anderson BA

[27] ↵

Appelle S
(1972) Perception and discrimination as a function of stimulus orientation: the 'oblique effect' in man and animals. Psychol Bull 78:266–278. doi:10.1037/h0033117 pmid:4562947
OpenUrl CrossRef PubMed

[29] Appelle S

[30] ↵

Bang JW,
Rahnev D
(2017) Stimulus expectation alters decision criterion but not sensory signal in perceptual decision making. Sci Rep 7:17072. doi:10.1038/s41598-017-16885-2 pmid:29213117
OpenUrl CrossRef PubMed

[32] Bang JW,

[33] Rahnev D

[34] ↵

Bays PM,
Catalao RF,
Husain M
(2009) The precision of visual working memory is set by allocation of a shared resource. J Vis 9:7.1. doi:10.1167/9.10.7 pmid:19810788
OpenUrl CrossRef PubMed

[36] Bays PM,

[37] Catalao RF,

[38] Husain M

[39] ↵

Berens P
(2009) CircStat: a MATLAB toolbox for circular statistics. J Stat Softw 31:1–21.
OpenUrl CrossRef PubMed

[41] Berens P

[42] ↵

Bergen JR,
Julesz B
(1983) Parallel versus serial processing in rapid pattern discrimination. Nature 303:696–698. doi:10.1038/303696a0 pmid:6855915
OpenUrl CrossRef PubMed

[44] Bergen JR,

[45] Julesz B

[46] ↵

Brainard DH,
Vision S
(1997) The psychophysics toolbox. Spatial vision 10:433–436.
OpenUrl CrossRef PubMed

[48] Brainard DH,

[49] Vision S

[50] ↵

Brouwer GJ,
Heeger DJ
(2009) Decoding and reconstructing color from responses in human visual cortex. J Neurosci 29:13992–14003. doi:10.1523/JNEUROSCI.3577-09.2009 pmid:19890009
OpenUrl Abstract/FREE Full Text

[52] Brouwer GJ,

[53] Heeger DJ

[54] ↵

Brunton BW,
Botvinick MM,
Brody CD
(2013) Rats and humans can optimally accumulate evidence for decision-making. Science 340:95–98. doi:10.1126/science.1233912 pmid:23559254
OpenUrl Abstract/FREE Full Text

[56] Brunton BW,

[57] Botvinick MM,

[58] Brody CD

[59] ↵

Chung E,
Romano JP
(2013) Exact and asymptotically robust permutation tests. Ann Stat 41:484–507.
OpenUrl

[61] Chung E,

[62] Romano JP

[63] ↵

Cortese A,
Amano K,
Koizumi A,
Kawato M,
Lau H
(2016) Multivoxel neurofeedback selectively modulates confidence without changing perceptual performance. Nat Commun 7:13669. doi:10.1038/ncomms13669 pmid:27976739
OpenUrl CrossRef PubMed

[65] Cortese A,

[66] Amano K,

[67] Koizumi A,

[68] Kawato M,

[69] Lau H

[70] ↵

deBettencourt MT,
Cohen JD,
Lee RF,
Norman KA,
Turk-Browne NB
(2015) Closed-loop training of attention with real-time brain imaging. Nat Neurosci 18:470–475. doi:10.1038/nn.3940 pmid:25664913
OpenUrl CrossRef PubMed

[72] deBettencourt MT,

[73] Cohen JD,

[74] Lee RF,

[75] Norman KA,

[76] Turk-Browne NB

[77] ↵

Demiralp T,
Bayraktaroglu Z,
Lenz D,
Junge S,
Busch NA,
Maess B,
Ergen M,
Herrmann CS
(2007) Gamma amplitudes are coupled to theta phase in human EEG during visual perception. Int J Psychophysiol 64:24–30. doi:10.1016/j.ijpsycho.2006.07.005 pmid:16956685
OpenUrl CrossRef PubMed

[79] Demiralp T,

[80] Bayraktaroglu Z,

[81] Lenz D,

[82] Junge S,

[83] Busch NA,

[84] Maess B,

[85] Ergen M,

[86] Herrmann CS

[87] ↵

Eldar E,
Bae GJ,
Kurth-Nelson Z,
Dayan P,
Dolan RJ
(2018) Magnetoencephalography decoding reveals structural differences within integrative decision processes. Nat Hum Behav 2:670–681. doi:10.1038/s41562-018-0423-3 pmid:31346283
OpenUrl CrossRef PubMed

[89] Eldar E,

[90] Bae GJ,

[91] Kurth-Nelson Z,

[92] Dayan P,

[93] Dolan RJ

[94] ↵

Emmanouil TA,
Treisman A
(2008) Dividing attention across feature dimensions in statistical processing of perceptual groups. Percept Psychophys 70:946–954. doi:10.3758/pp.70.6.946 pmid:18717382
OpenUrl CrossRef PubMed

[96] Emmanouil TA,

[97] Treisman A

[98] ↵

Failing M,
Theeuwes J
(2018) Selection history: how reward modulates selectivity of visual attention. Psychon Bull Rev 25:514–538. doi:10.3758/s13423-017-1380-y pmid:28986770
OpenUrl CrossRef PubMed

[100] Failing M,

[101] Theeuwes J

[102] ↵

Fetsch CR,
DeAngelis GC,
Angelaki DE
(2013) Bridging the gap between theories of sensory cue integration and the physiology of multisensory neurons. Nat Rev Neurosci 14:429–442. doi:10.1038/nrn3503 pmid:23686172
OpenUrl CrossRef PubMed

[104] Fetsch CR,

[105] DeAngelis GC,

[106] Angelaki DE

[107] ↵

Gao J,
Tortell R,
McClelland JL
(2011) Dynamic integration of reward and stimulus information in perceptual decision-making. PLoS One 6:e16749. doi:10.1371/journal.pone.0016749 pmid:21390225
OpenUrl CrossRef PubMed

[109] Gao J,

[110] Tortell R,

[111] McClelland JL

[112] ↵

Garcia JO,
Srinivasan R,
Serences JT
(2013) Near-real-time feature-selective modulations in human cortex. Curr Biol 23:515–522. doi:10.1016/j.cub.2013.02.013 pmid:23477721
OpenUrl CrossRef PubMed

[114] Garcia JO,

[115] Srinivasan R,

[116] Serences JT

[117] ↵

Gold JI,
Shadlen MN
(2007) The neural basis of decision making. Annu Rev Neurosci 30:535–574. doi:10.1146/annurev.neuro.29.051605.113038 pmid:17600525
OpenUrl CrossRef PubMed

[119] Gold JI,

[120] Shadlen MN

[121] ↵

Habes I,
Rushton S,
Johnston SJ,
Sokunbi MO,
Barawi K,
Brosnan M,
Daly T,
Ihssen N,
Linden DEJ
(2016) fMRI neurofeedback of higher visual areas and perceptual biases. Neuropsychologia 85:208–215. doi:10.1016/j.neuropsychologia.2016.03.031 pmid:27020139
OpenUrl CrossRef PubMed

[123] Habes I,

[124] Rushton S,

[125] Johnston SJ,

[126] Sokunbi MO,

[127] Barawi K,

[128] Brosnan M,

[129] Daly T,

[130] Ihssen N,

[131] Linden DEJ

[132] ↵

Ho T,
Brown S,
van Maanen L,
Forstmann BU,
Wagenmakers EJ,
Serences JT
(2012) The optimality of sensory processing during the speed–accuracy tradeoff. J Neurosci 32:7992–8003. doi:10.1523/JNEUROSCI.0340-12.2012 pmid:22674274
OpenUrl Abstract/FREE Full Text

[134] Ho T,

[135] Brown S,

[136] van Maanen L,

[137] Forstmann BU,

[138] Wagenmakers EJ,

[139] Serences JT

[140] ↵

John-Saaltink ES,
Kok P,
Lau HC,
de Lange FP
(2016) Serial dependence in perceptual decisions is reflected in activity patterns in primary visual cortex. J Neurosci 36:6186–6192. doi:10.1523/JNEUROSCI.4390-15.2016 pmid:27277797
OpenUrl Abstract/FREE Full Text

[142] John-Saaltink ES,

[143] Kok P,

[144] Lau HC,

[145] de Lange FP

[146] ↵

Kelly SP,
O'Connell RG
(2015) The neural processes underlying perceptual decision making in humans: recent progress and future directions. J Physiol Paris 109:27–37. doi:10.1016/j.jphysparis.2014.08.003 pmid:25204272
OpenUrl CrossRef PubMed

[148] Kelly SP,

[149] O'Connell RG

[150] ↵

Krekelberg B
(2021) klabhub/bayesFactor: version 2.1.0. dding log numerics and linear model analyses. Zenodo. Available at https://zenodo.org/record/4849568.

[152] Krekelberg B

[153] ↵

Landy MS,
Maloney LT,
Johnston EB,
Young M
(1995) Measurement and modeling of depth cue combination: in defense of weak fusion. Vision Res 35:389–412. doi:10.1016/0042-6989(94)00176-m pmid:7892735
OpenUrl CrossRef PubMed

[155] Landy MS,

[156] Maloney LT,

[157] Johnston EB,

[158] Young M

[159] ↵

Ma WJ
(2010) Signal detection theory, uncertainty, and Poisson-like population codes. Vision Res 50:2308–2319. doi:10.1016/j.visres.2010.08.035 pmid:20828581
OpenUrl CrossRef PubMed

[161] Ma WJ

[162] ↵

Maunsell JH,
Treue S
(2006) Feature-based attention in visual cortex. Trends Neurosci 29:317–322. doi:10.1016/j.tins.2006.04.001 pmid:16697058
OpenUrl CrossRef PubMed

[164] Maunsell JH,

[165] Treue S

[166] ↵

Morey RD,
Rouder JN
(2011) Bayes factor approaches for testing interval null hypotheses. Psychol Methods 16:406–419. doi:10.1037/a0024377 pmid:21787084
OpenUrl CrossRef PubMed

[168] Morey RD,

[169] Rouder JN

[170] ↵

Mostert P,
Kok P,
de Lange FP
(2015) Dissociating sensory from decision processes in human perceptual decision making. Sci Rep 5:18253. doi:10.1038/srep18253 pmid:26666393
OpenUrl CrossRef PubMed

[172] Mostert P,

[173] Kok P,

[174] de Lange FP

[175] ↵

Moxon KA,
Foffani G
(2015) Brain-machine interfaces beyond neuroprosthetics. Neuron 86:55–67. doi:10.1016/j.neuron.2015.03.036 pmid:25856486
OpenUrl CrossRef PubMed

[177] Moxon KA,

[178] Foffani G

[179] ↵

Müller MM,
Andersen S,
Trujillo NJ,
Valdés-Sosa P,
Malinowski P,
Hillyard SA
(2006) Feature-selective attention enhances color signals in early visual areas of the human brain. Proc Natl Acad Sci USA 103:14250–14254. doi:10.1073/pnas.0606668103 pmid:16956975
OpenUrl Abstract/FREE Full Text

[181] Müller MM,

[182] Andersen S,

[183] Trujillo NJ,

[184] Valdés-Sosa P,

[185] Malinowski P,

[186] Hillyard SA

[187] ↵

Nan W,
Wan F,
Lou CI,
Vai MI,
Rosa A
(2013) Peripheral visual performance enhancement by neurofeedback training. Appl Psychophysiol Biofeedback 38:285–291. doi:10.1007/s10484-013-9233-6 pmid:24101183
OpenUrl CrossRef PubMed

[189] Nan W,

[190] Wan F,

[191] Lou CI,

[192] Vai MI,

[193] Rosa A

[194] ↵

Norcia AM,
Appelbaum LG,
Ales JM,
Cottereau BR,
Rossion B
(2015) The steady-state visual evoked potential in vision research: a review. J Vis 15:4. doi:10.1167/15.6.4 pmid:26024451
OpenUrl Abstract/FREE Full Text

[196] Norcia AM,

[197] Appelbaum LG,

[198] Ales JM,

[199] Cottereau BR,

[200] Rossion B

[201] ↵

Oostenveld R,
Praamstra P
(2001) The five percent electrode system for high-resolution EEG and ERP measurements. Clin Neurophysiol 112:713–719. doi:10.1016/s1388-2457(00)00527-7 pmid:11275545
OpenUrl CrossRef PubMed

[203] Oostenveld R,

[204] Praamstra P

[205] ↵

Oostenveld R,
Fries P,
Maris E,
Schoffelen JM
(2011) FieldTrip: Open Source software for advanced analysis of MEG, EEG, and invasive electrophysiological data. Comput Intell Neurosci 2011:156869. doi:10.1155/2011/156869 pmid:21253357
OpenUrl CrossRef PubMed

[207] Oostenveld R,

[208] Fries P,

[209] Maris E,

[210] Schoffelen JM

[211] ↵

Painter DR,
Dux PE,
Travis SL,
Mattingley JB
(2014) Neural responses to target features outside a search array are enhanced during conjunction but not unique-feature search. J Neurosci 34:3390–3401. doi:10.1523/JNEUROSCI.3630-13.2014 pmid:24573295
OpenUrl Abstract/FREE Full Text

[213] Painter DR,

[214] Dux PE,

[215] Travis SL,

[216] Mattingley JB

[217] ↵

Purcell BA,
Heitz RP,
Cohen JY,
Schall JD,
Logan GD,
Palmeri TJ
(2010) Neurally constrained modeling of perceptual decision making. Psychol Rev 117:1113–1143. doi:10.1037/a0020311 pmid:20822291
OpenUrl CrossRef PubMed

[219] Purcell BA,

[220] Heitz RP,

[221] Cohen JY,

[222] Schall JD,

[223] Logan GD,

[224] Palmeri TJ

[225] ↵

Rahnev D,
Denison RN
(2018) Suboptimality in perceptual decision making. Behav Brain Sci 41:1–107. doi:10.1017/S0140525X18000936 pmid:29485020
OpenUrl CrossRef PubMed

[227] Rahnev D,

[228] Denison RN

[229] ↵

Rajagovindan R,
Ding M
(2011) From prestimulus alpha oscillation to visual-evoked response: an inverted-U function and its attentional modulation. J Cogn Neurosci 23:1379–1394. doi:10.1162/jocn.2010.21478 pmid:20459310
OpenUrl CrossRef PubMed

[231] Rajagovindan R,

[232] Ding M

[233] ↵

Rangelov D,
West R,
Mattingley JB
(2020) Stimulus reliability automatically biases temporal integration of discrete perceptual targets. bioRxiv 841353.

[235] Rangelov D,

[236] West R,

[237] Mattingley JB

[238] ↵

Raposo D,
Sheppard JP,
Schrater PR,
Churchland AK
(2012) Multisensory decision-making in rats and humans. J Neurosci 32:3726–3735. doi:10.1523/JNEUROSCI.4998-11.2012 pmid:22423093
OpenUrl Abstract/FREE Full Text

[240] Raposo D,

[241] Sheppard JP,

[242] Schrater PR,

[243] Churchland AK

[244] ↵

Ratcliff R,
McKoon G
(2008) The diffusion decision model: theory and data for two-choice decision tasks. Neural Comput 20:873–922. doi:10.1162/neco.2008.12-06-420 pmid:18085991
OpenUrl CrossRef PubMed

[246] Ratcliff R,

[247] McKoon G

[248] ↵

Ratcliff R,
Rouder JN
(1998) Modeling response times for two-choice decisions. Psychol Sci 9:347–356. doi:10.1111/1467-9280.00067
OpenUrl CrossRef PubMed

[250] Ratcliff R,

[251] Rouder JN

[252] ↵

Renton AI,
Mattingley JB,
Painter DR
(2019a) Optimising non-invasive brain-computer interface systems for free communication between naïve human participants. Sci Rep 9:18705. doi:10.1038/s41598-019-55166-y pmid:31822715
OpenUrl CrossRef PubMed

[254] Renton AI,

[255] Mattingley JB,

[256] Painter DR

[257] ↵

Renton AI,
Painter DR,
Mattingley JB
(2019b) Differential deployment of visual attention during interactive approach and avoidance behavior. Cereb Cortex 29:2366–2383. doi:10.1093/cercor/bhy105 pmid:29750259
OpenUrl CrossRef PubMed

[259] Renton AI,

[260] Painter DR,

[261] Mattingley JB

[262] ↵

Robineau F,
Rieger SW,
Mermoud C,
Pichon S,
Koush Y,
Van De Ville D,
Vuilleumier P,
Scharnowski F
(2014) Self-regulation of inter-hemispheric visual cortex balance through real-time fMRI neurofeedback training. Neuroimage 100:1–14. doi:10.1016/j.neuroimage.2014.05.072 pmid:24904993
OpenUrl CrossRef PubMed

[264] Robineau F,

[265] Rieger SW,

[266] Mermoud C,

[267] Pichon S,

[268] Koush Y,

[269] Van De Ville D,

[270] Vuilleumier P,

[271] Scharnowski F

[272] ↵

Rouder JN,
Speckman PL,
Sun D,
Morey RD,
Iverson G
(2009) Bayesian t tests for accepting and rejecting the null hypothesis. Psychon Bull Rev 16:225–237. doi:10.3758/PBR.16.2.225 pmid:19293088
OpenUrl CrossRef PubMed

[274] Rouder JN,

[275] Speckman PL,

[276] Sun D,

[277] Morey RD,

[278] Iverson G

[279] ↵

Saenz M,
Buracas GT,
Boynton GM
(2002) Global effects of feature-based attention in human visual cortex. Nat Neurosci 5:631–632. doi:10.1038/nn876 pmid:12068304
OpenUrl CrossRef PubMed

[281] Saenz M,

[282] Buracas GT,

[283] Boynton GM

[284] ↵

Salzman CD,
Britten KH,
Newsome WT
(1990) Cortical microstimulation influences perceptual judgments of motion direction. Nature 346:174–177. doi:10.1038/346174a0 pmid:2366872
OpenUrl CrossRef PubMed

[286] Salzman CD,

[287] Britten KH,

[288] Newsome WT

[289] ↵

Scharnowski F,
Hutton C,
Josephs O,
Weiskopf N,
Rees G
(2012) Improving visual perception through neurofeedback. J Neurosci 32:17830–17841. doi:10.1523/JNEUROSCI.6334-11.2012 pmid:23223302
OpenUrl Abstract/FREE Full Text

[291] Scharnowski F,

[292] Hutton C,

[293] Josephs O,

[294] Weiskopf N,

[295] Rees G

[296] ↵

Schneegans S,
Bays PM
(2016) No fixed item limit in visuospatial working memory. Cortex 83:181–193. doi:10.1016/j.cortex.2016.07.021 pmid:27565636
OpenUrl CrossRef PubMed

[298] Schneegans S,

[299] Bays PM

[300] ↵

Scolari M,
Byers A,
Serences JT
(2012) Optimal deployment of attentional gain during fine discriminations. J Neurosci 32:7723–7733. doi:10.1523/JNEUROSCI.5558-11.2012 pmid:22649250
OpenUrl Abstract/FREE Full Text

[302] Scolari M,

[303] Byers A,

[304] Serences JT

[305] ↵

Shibata K,
Watanabe T,
Kawato M,
Sasaki Y
(2016) Differential activation patterns in the same brain region led to opposite emotional states. PLoS Biol 14:e1002546. doi:10.1371/journal.pbio.1002546 pmid:27608359
OpenUrl CrossRef PubMed

[307] Shibata K,

[308] Watanabe T,

[309] Kawato M,

[310] Sasaki Y

[311] ↵

Shibata K,
Watanabe T,
Sasaki Y,
Kawato M
(2011) Perceptual learning incepted by decoded fMRI neurofeedback without stimulus presentation. Science 334:1413–1415. doi:10.1126/science.1212003 pmid:22158821
OpenUrl Abstract/FREE Full Text

[313] Shibata K,

[314] Watanabe T,

[315] Sasaki Y,

[316] Kawato M

[317] ↵

Sitaram R,
Ros T,
Stoeckel L,
Haller S,
Scharnowski F,
Lewis-Peacock J,
Weiskopf N,
Blefari ML,
Rana M,
Oblak E,
Birbaumer N,
Sulzer J
(2017) Closed-loop brain training: the science of neurofeedback. Nat Rev Neurosci 18:86–100. doi:10.1038/nrn.2016.164 pmid:28003656
OpenUrl CrossRef PubMed

[319] Sitaram R,

[320] Ros T,

[321] Stoeckel L,

[322] Haller S,

[323] Scharnowski F,

[324] Lewis-Peacock J,

[325] Weiskopf N,

[326] Blefari ML,

[327] Rana M,

[328] Oblak E,

[329] Birbaumer N,

[330] Sulzer J

[331] ↵

Smout CA,
Tang MF,
Garrido MI,
Mattingley JB
(2019) Attention promotes the neural encoding of prediction errors. PLoS Biol 17:e2006812. doi:10.1371/journal.pbio.2006812 pmid:30811381
OpenUrl CrossRef PubMed

[333] Smout CA,

[334] Tang MF,

[335] Garrido MI,

[336] Mattingley JB

[337] ↵

Sorger B,
Scharnowski F,
Linden DE,
Hampson M,
Young KD
(2019) Control freaks: towards optimal selection of control conditions for fMRI neurofeedback studies. Neuroimage 186:256–265. doi:10.1016/j.neuroimage.2018.11.004 pmid:30423429
OpenUrl CrossRef PubMed

[339] Sorger B,

[340] Scharnowski F,

[341] Linden DE,

[342] Hampson M,

[343] Young KD

[344] ↵

Spitzer B,
Waschke L,
Summerfield C
(2017) Selective overweighting of larger magnitudes during noisy numerical comparison. Nat Hum Behav 1:145–148. doi:10.1038/s41562-017-0145 pmid:32340412
OpenUrl CrossRef PubMed

[346] Spitzer B,

[347] Waschke L,

[348] Summerfield C

[349] ↵

Strehl U
(2014) What learning theories can teach us in designing neurofeedback treatments. Front Hum Neurosci 8:894. doi:10.3389/fnhum.2014.00894 pmid:25414659
OpenUrl CrossRef PubMed

[351] Strehl U

[352] ↵

Summerfield C,
De Lange FP
(2014) Expectation in perceptual decision making: neural and computational mechanisms. Nat Rev Neurosci 15:745–756. doi:10.1038/nrn3838 pmid:25315388
OpenUrl CrossRef PubMed

[354] Summerfield C,

[355] De Lange FP

[356] ↵

Summerfield C,
Tsetsos K
(2015) Do humans make good decisions? Trends Cogn Sci 19:27–34. doi:10.1016/j.tics.2014.11.005
OpenUrl CrossRef PubMed

[358] Summerfield C,

[359] Tsetsos K

[360] ↵

Tadel F,
Baillet S,
Mosher JC,
Pantazis D,
Leahy RM
(2011) Brainstorm: a user-friendly application for MEG/EEG analysis. Comput Intell Neurosci 2011:879716. doi:10.1155/2011/879716 pmid:21584256
OpenUrl CrossRef PubMed

[362] Tadel F,

[363] Baillet S,

[364] Mosher JC,

[365] Pantazis D,

[366] Leahy RM

[367] ↵

Tang MF,
Ford L,
Arabzadeh E,
Enns JT,
Visser TA,
Mattingley JB
(2020) Neural dynamics of the attentional blink revealed by encoding orientation selectivity during rapid visual presentation. Nat Commun 11:434. doi:10.1038/s41467-019-14107-z pmid:31974370
OpenUrl CrossRef PubMed

[369] Tang MF,

[370] Ford L,

[371] Arabzadeh E,

[372] Enns JT,

[373] Visser TA,

[374] Mattingley JB

[375] ↵

Tang MF,
Smout CA,
Arabzadeh E,
Mattingley JB
(2018) Prediction error and repetition suppression have distinct effects on neural representations of visual information. eLife 7:e33123. doi:10.7554/eLife.33123
OpenUrl CrossRef PubMed

[377] Tang MF,

[378] Smout CA,

[379] Arabzadeh E,

[380] Mattingley JB

[381] ↵

Theeuwes J,
Olivers CNL,
Belopolsky A
(2010) Stimulus-driven capture and contingent capture. Wiley Interdiscip Rev Cogn Sci 1:872–881. doi:10.1002/wcs.83 pmid:26271783
OpenUrl CrossRef PubMed

[383] Theeuwes J,

[384] Olivers CNL,

[385] Belopolsky A

[386] ↵

Treisman AM
(1969) Strategies and models of selective attention. Psychol Rev 76:282–299. doi:10.1037/h0027242 pmid:4893203
OpenUrl CrossRef PubMed

[388] Treisman AM

[389] ↵

Treue S,
Trujillo JC
(1999) Feature-based attention influences motion processing gain in macaque visual cortex. Nature 399:575–579. doi:10.1038/21176 pmid:10376597
OpenUrl CrossRef PubMed

[391] Treue S,

[392] Trujillo JC

[393] ↵

Trommershäuser J
(2009) Biases and optimality of sensory-motor and cognitive decisions. Prog Brain Res 174:267–278. doi:10.1016/S0079-6123(09)01321-1 pmid:19477345
OpenUrl CrossRef PubMed

[395] Trommershäuser J

[396] ↵

Tsetsos K,
Gao J,
McClelland JL,
Usher M
(2012) Using time-varying evidence to test models of decision dynamics: bounded diffusion vs. the leaky competing accumulator model. Front Neurosci 6:79. doi:10.3389/fnins.2012.00079 pmid:22701399
OpenUrl CrossRef PubMed

[398] Tsetsos K,

[399] Gao J,

[400] McClelland JL,

[401] Usher M

[402] ↵

Usher M,
McClelland JL
(2001) The time course of perceptual choice: the leaky, competing accumulator model. Psychol Rev 108:550–592. doi:10.1037/0033-295x.108.3.550 pmid:11488378
OpenUrl CrossRef PubMed

[404] Usher M,

[405] McClelland JL

[406] ↵

van Vugt MK,
Beulen MA,
Taatgen NA
(2019) Relation between centro-parietal positivity and diffusion model parameters in both perceptual and memory-based decision making. Brain Res 1715:1–12. doi:10.1016/j.brainres.2019.03.008 pmid:30876858
OpenUrl CrossRef PubMed

[408] van Vugt MK,

[409] Beulen MA,

[410] Taatgen NA

[411] ↵

Wan F,
da Cruz JN,
Nan W,
Wong CM,
Vai MI,
Rosa A
(2016) Alpha neurofeedback training improves SSVEP-based BCI performance. J Neural Eng 13:036019. doi:10.1088/1741-2560/13/3/036019
OpenUrl CrossRef

[413] Wan F,

[414] da Cruz JN,

[415] Nan W,

[416] Wong CM,

[417] Vai MI,

[418] Rosa A

[419] ↵

Waskom ML,
Asfour J,
Kiani R
(2018) Perceptual insensitivity to higher-order statistical moments of coherent random dot motion. J Vis 18:9. doi:10.1167/18.6.9 pmid:30029220
OpenUrl CrossRef PubMed

[421] Waskom ML,

[422] Asfour J,

[423] Kiani R

[424] ↵

Watanabe T,
Sasaki Y,
Shibata K,
Kawato M
(2017) Advances in fMRI real-time neurofeedback. Trends Cogn Sci 21:997–1010. doi:10.1016/j.tics.2017.09.010 pmid:29031663
OpenUrl CrossRef PubMed

[426] Watanabe T,

[427] Sasaki Y,

[428] Shibata K,

[429] Kawato M

[430] ↵

Wolfe JM,
Horowitz TS
(2004) What attributes guide the deployment of visual attention and how do they do it? Nat Rev Neurosci 5:495–501. doi:10.1038/nrn1411
OpenUrl CrossRef PubMed

[432] Wolfe JM,

[433] Horowitz TS

[434] ↵

Yantis S
(1993) Stimulus-driven attentional capture. Curr Dir Psychol Sci 2:156–161. doi:10.1111/1467-8721.ep10768973
OpenUrl CrossRef

[436] Yantis S

[437] ↵

Zhang D,
Maye A,
Gao X,
Hong B,
Engel AK,
Gao S
(2010) An independent brain–computer interface using covert non-spatial visual selective attention. J Neural Eng 7:16010. doi:10.1088/1741-2560/7/1/016010 pmid:20083864
OpenUrl CrossRef PubMed

[439] Zhang D,

[440] Maye A,

[441] Gao X,

[442] Hong B,

[443] Engel AK,

[444] Gao S

[445] ↵

Zylberberg A,
Fetsch CR,
Shadlen MN
(2016) The influence of evidence volatility on choice, reaction time and confidence in a perceptual decision. eLife 5:e17688. doi:10.7554/eLife.17688
OpenUrl CrossRef PubMed

[447] Zylberberg A,

[448] Fetsch CR,

[449] Shadlen MN

Main menu

User menu

Search

Implicit Neurofeedback Training of Feature-Based Attention Promotes Biased Sensory Processing during Integrative Decision-Making

Abstract

Introduction

Materials and Methods

Participants

Task overview

Experiment protocol

Individual trial design

Block design and counterbalancing

Frequency tagging design

Stimulus design

Display computer specifications

EEG recording and processing

EEG recording

Neurofeedback

Motion coherence

Real-time SSVEP selectivity

Staircase procedure

EEG preprocessing for offline analyses

Statistical tests and models

Permutation tests

Forward encoding model

Mixture distribution modeling of behavioral responses

Source code and data availability

Results

Implicit training: debriefing questionnaire

Sensory processing biases: neurofeedback training of attention

Frequency tagging

SSVEP selectivity during test trials

SSVEP selectivity during neurofeedback training

Integrative decision-making processes: neural representations of the trained, untrained, and average motion directions

Forward encoding model

Forward encoding results: trained and untrained motion directions

Forward encoding results: average motion direction

Integrative decision-making outcomes: behavioral reporting of the average motion

Response bias distribution

Feature decision weightings

Response bias by attentional selectivity

Neurofeedback dynamics

Feedback variability

Learning curves and switch costs

Discussion

Footnotes

References

In this issue

Citation Manager Formats

Jump to section

Keywords

Responses to this article

Jump to comment:

Related Articles

Cited By...

More in this TOC Section

Research Articles

Behavioral/Cognitive