Recent evidence suggests that voluntary spatial attention can affect neural processing of visual stimuli that do not enter conscious awareness (i.e., invisible stimuli), supporting the notion that attention and awareness are dissociable processes [Wyart, V., Dehaene, S., & Tallon-Baudry, C. Early dissociation between neural signatures of endogenous spatial attention and perceptual awareness during visual masking. Frontiers in Human Neuroscience, 6, 1–14, 2012; Watanabe, M., Cheng, K., Murayama, Y., Ueno, K., Asamizuya, T., Tanaka, K., et al. Attention but not awareness modulates the BOLD signal in the human V1 during binocular suppression. Science, 334, 829–831, 2011]. To date, however, no study has demonstrated that these effects reflect enhancement of the neural representation of invisible stimuli per se, as opposed to other neural processes not specifically tied to the stimulus in question. In addition, it remains unclear whether spatial attention can modulate neural representations of invisible stimuli in direct competition with highly salient and visible stimuli. Here we developed a novel EEG frequency-tagging paradigm to obtain a continuous readout of human brain activity associated with visible and invisible signals embedded in dynamic noise. Participants (n = 23) detected occasional contrast changes in one of two flickering image streams on either side of fixation. Each image stream contained a visible or invisible signal embedded in every second noise image, the visibility of which was titrated and checked using a two-interval forced-choice detection task. Steady-state visual-evoked potentials were computed from EEG data at the signal and noise frequencies of interest. Cluster-based permutation analyses revealed significant neural responses to both visible and invisible signals across posterior scalp electrodes. Control analyses revealed that these responses did not reflect a subharmonic response to noise stimuli. In line with previous findings, spatial attention increased the neural representation of visible signals. Crucially, spatial attention also increased the neural representation of invisible signals. As such, the present results replicate and extend previous studies by demonstrating that attention can modulate the neural representation of invisible signals that are in direct competition with highly salient masking stimuli.
When viewing a cluttered visual scene, representations of the various objects compete for limited neural resources (Desimone & Duncan, 1995; Broadbent, 1958). Such ongoing neural competition can be biased by top–down mechanisms to facilitate the observer's behavioral goals (Beck & Kastner, 2009). For example, voluntarily allocating covert spatial attention to a specific region of the visual field can selectively boost neural representations of salient stimuli within that region (Martinez et al., 1999; Hillyard & Anllo-Vento, 1998; Müller et al., 1998). Interestingly, recent studies demonstrate that spatial attention can also affect neural processing of weak stimuli that do not enter awareness (equated here with the contents of conscious experience; Wyart, Dehaene, & Tallon-Baudry, 2012; Schurger, Cowey, Cohen, Treisman, & Tallon-Baudry, 2008; Wyart & Tallon-Baudry, 2008). However, because attention encompasses a variety of neural mechanisms (for a review, see Womelsdorf & Everling, 2015), it remains unclear which subcomponents activate during processing of invisible stimuli. In particular, no study to date has tied neural activity to specific invisible stimuli, and thus it remains unclear whether spatial attention enhances neural representations of invisible stimuli or merely activates other neural mechanisms not specific to neural representations per se (e.g., alerting, orienting, or suppression mechanisms). Evidence that spatial attention increases the neural representation of invisible stimuli without a corresponding increase in object awareness would provide clear evidence that attention and awareness dissociate at the level of stimulus representations. Furthermore, previous studies presented invisible stimuli at different times or locations to highly visible masking stimuli, and thus, it remains unclear how spatial attention treats neural representations of invisible signals that are in direct competition with visible stimuli. Such research is necessary if we are to understand how top–down mechanisms in the visual system allocate limited resources to competing stimuli with different levels of bottom–up signal strength (i.e., salience). In this study, we used EEG to measure neural representations of visible and invisible stimuli embedded in highly salient noise and assessed the effect of voluntary covert spatial attention on these neural representations.
To investigate these questions, it is necessary to disambiguate relatively weak neural activity arising from subjectively invisible targets from the stronger responses associated with highly salient and spatially coincident masking stimuli. To date, however, no such technique has been devised to effectively distinguish the neural signatures of these weak and strong sensory inputs. If a train of stimuli is presented at a fixed frequency, however, a stable oscillatory response is produced in the brain that can be observed in the frequency domain in EEG recordings (the steady-state visual-evoked potential [SSVEP]; Regan, 1966). Multiple stimuli in a visual scene can thus be “frequency tagged” when flickered at unique frequencies, an approach that has proven useful for exploring the effects of attention on visible stimuli at separate spatial locations (Norcia, Appelbaum, Ales, Cottereau, & Rossion, 2015).
A recent study by Ales, Farzin, Rossion, and Norcia (2012) pioneered a novel SSVEP technique for measuring neural representations of signals embedded in dynamic noise. In their study, Ales et al. presented participants with streams of luminance- and amplitude-matched noise images at a rate of 6 Hz. Every second image contained a face stimulus embedded in noise, and the coherence of the face was gradually increased over the duration of the trial until participants indicated they had detected it. Crucially, power at the frequency of signal presentation (3 Hz, representing the face in every second image) was found only in trials that contained embedded faces and not in trials in which the face was replaced by another noise display. Thus, the neural activity at the frequency of the embedded signal serves as a useful measure of the neural representation of that stimulus, irrespective of any other neural processes that may be operating concurrently.
Using the same principle as Ales et al. (2012), we developed a novel paradigm to obtain a continuous readout of neural activity associated with visible and invisible signals embedded in dynamic noise. Participants directed attention to one of a pair of flickering image streams to detect occasional contrast changes, and we assessed the effect of spatial attention on neural representations of both visible and invisible signals. We employed a two-interval, forced-choice signal detection task to confirm that appropriate levels of signal coherence were selected for visible and invisible signals. To anticipate, we found that spatial attention enhanced neural representations of both visible and invisible signals, suggesting that attention can bias neural activity in favor of invisible stimuli that are in spatial and temporal competition with highly salient masking noise.
Twenty-three healthy participants (11 women, mean age = 22.65 years) with normal or corrected-to-normal vision were recruited via an online research participation scheme at The University of Queensland. Participants completed a safety screening questionnaire and provided written informed consent before commencement of the study, which was approved by The University of Queensland human research ethics committee.
Stimuli and Apparatus
The method of stimulus generation (Figure 1) was adapted from Ales et al. (2012) to maintain the same average power distribution and luminance across all images. All images were created from the same seed image consisting of an annulus (seven cycles, inner diameter = 4.67° of visual angle, outer diameter = 14° of visual angle) on a uniform midgray square background (14° of visual angle; Figure 1A, top left). The phase distribution of the seed image was randomized to create a noise background with the same amplitude distribution as the seed image (Figure 1A, bottom left). The annulus and noise background were then combined using complementary spatial blending masks (which spanned from the annulus edges to 2° of visual angle within each edge; Figure 1A, top and bottom right) to create an exemplar image consisting of a fully coherent annulus on a noise background (Figure 1A, center right). The phase distribution of this exemplar image was then “scrambled” (randomized) to the extent required by the trial sequence (see Stimulation Protocol, below): phase angles of “noise images” were scrambled completely (Figure 1B, bottom), whereas phase angles of “signal images” were linearly interpolated between the original phase angles and a random phase distribution (Figure 1B, top and middle). Because phase angles are circular, interpolation of phase angles was computed in the direction of least difference to maintain a uniform phase distribution (Ales et al., 2012).
Thus, all images contained some amount of “noise,” which represented the (partially or completely) randomized phase distribution of its exemplar image. Signal images contained noise both “behind” the annulus (in the exemplar background), as well as “in front of” the annulus, because the phase structure of the exemplar image was partially randomized in the final image creation step. Because each exemplar image was created using a unique noise background, the only consistent structure between any two images was the signal itself, subject to its level of phase coherence. Furthermore, because all images—both signal and noise—were created from the same seed annulus image, all images in the experiment shared the same low-level characteristics, including amplitude and luminance.
Stimuli were presented on a 21-in. CRT monitor (NEC, Accusync 120) with a screen resolution of 800 × 600 pixels and a refresh rate of 120 Hz, using the Cogent 2000 Toolbox (www.vislab.ucl.ac.uk/cogent.php) for MATLAB (The Mathworks, Inc.) running under Windows XP. Participants were seated in a comfortable armchair in an electrically shielded laboratory, with the head supported by a chin rest at a viewing distance of 57 cm.
This study used a within-participant design with two levels of target awareness (visible, invisible) and two levels of spatial attention (attended, ignored). Two tasks with similar overall designs were employed to manipulate awareness and spatial attention.
Participants were presented with two flickering image streams on either side of fixation (visual angle = 14°), as shown in Figure 2 and Movie 1. Each image stream contained two consecutive intervals of 2.4-sec duration (see Stimulation Protocol for interval details). One of the intervals in each image stream (randomized separately) contained signal (the “signal interval”), and the other interval contained noise only (the “noise interval”). Participants were asked to maintain fixation and report, on the cued side, which of the two intervals contained signal (two-interval forced-choice) while ignoring the noncued side. The cue direction (left or right) was randomized for the first trial of each block and then alternated every eight trials.
Participants completed two versions of the Awareness Task. The first version was run at the beginning of the experiment (following practice with accuracy feedback) to set signal coherence levels for the subsequent Attention Task (see below). In this version, participants completed one block of 48 trials (with feedback), during which levels of signal coherence were adjusted according to an adaptive Quest staircase (Watson & Pelli, 1983) designed to approximate the maximum level of signal coherence that could not be detected by each participant (i.e., the invisible condition). Signal coherence for the visible condition was then set 40% higher than this level, as guided by psychometric functions fitted to pilot data. The second version of the Awareness Task was run at the end of the experiment to verify that appropriate levels of signal coherence had been selected. In this version, participants completed two blocks of 64 trials (without feedback), with each image stream containing visible or invisible signal in one of the two consecutive intervals (randomized separately across trials).
Participants were again presented with two flickering image streams on either side of fixation, as shown in Figure 3 and Movie 2. Unlike in the Awareness Task, however, in the Attention Task each image stream contained only one interval of 10-sec duration per trial, and both image streams contained either visible or invisible signals (as per the staircase procedure above). In addition, each image stream occasionally decreased in contrast before returning to normal across a 1-sec period (ramping on and off linearly), with at least 1.5 sec between peaks of contrast decreases (in either stream). Participants were asked to maintain fixation and report at the end of the trial how many contrast decreases (targets) occurred in the cued (attended) image stream. When the attended stream contained two contrast targets, the second target peaked between 7 and 8.5 sec into the trial to encourage sustained attention throughout trials. Participants were allowed to practice the task (with feedback after each trial) before completing eight blocks of 64 test trials, with feedback provided between blocks. The percentage of contrast decrease was adjusted between blocks to maintain an approximate detection level of 65% (according to a 1 up/2 down staircase with step sizes of 5%).
During any one trial, intervals in the left and right image streams flickered at unique frequencies (10 and 15 Hz, counterbalanced across trials). Although Awareness Task trials contained two intervals per image stream and Attention Task trials contained only one interval per image stream, the structure of intervals in both tasks was essentially the same. Figure 4 shows the stimulation protocol for one interval flickering at 10 Hz. All intervals began with 0.5 sec of static noise, after which images flickered consecutively at the designated frequency (10 or 15 Hz). The phase distributions of all images in “noise intervals” (Awareness Task) were completely scrambled (see Stimuli and Apparatus). During “signal intervals” (Awareness and Attention Tasks), images alternated between completely phase-scrambled images (noise) and partially phase-scrambled images (signal). The coherence of signal images ramped up linearly during the first 0.4 sec of signal intervals to eliminate involuntary capture of attention (Figure 4). At the end of the flicker duration (2.4 sec for Awareness Task trials, 10.4 sec for Attention Task trials), static noise was presented until the next interval began flickering (first interval of Awareness Task trials only) or the participant responded (Attention Task trials and second interval of Awareness Task trials).
Shown at the top of Figure 4 are putative neural responses evoked by the stimulation protocol. Because all flickering images contained some amount of “noise,” SSVEP responses were expected to be elicited by noise stimuli at the noise frequency (i.e., 10 or 15 Hz). Crucially, because a signal was embedded in every second image during signal intervals, a separate SSVEP was expected to be elicited at half the noise frequency in response to signal (5 or 7.5 Hz, the signal frequency). Thus, we were able to isolate neural responses to both noise and signal (at two levels of awareness) when those stimuli were either attended or ignored (see Results for details of power computation).
Participants were fitted with a 64 Ag-AgCl electrode EEG system (BioSemi Active Two) after the initial Awareness Task, and EEG data were recorded during the Attention Task and final Awareness Task. Continuous data were recorded using BioSemi ActiView software (www.biosemi.com) and were digitized at a sample rate of 1024 Hz with 24-bit A/D conversion and a 0.01–208 Hz amplifier band pass. All scalp channels were referenced to the standard BioSemi reference electrodes, and electrode offsets were adjusted to be below 25 μV before beginning the recording. Horizontal and vertical eye movements were recorded via pairs of BioSemi flat Ag-AgCl electrooculographic electrodes placed to the outside of each eye, and above and below the left eye, respectively.
EEG Data Preprocessing
EEG recordings were processed offline using the Fieldtrip toolbox in MATLAB (Oostenveld, Fries, Maris, & Schoffelen, 2011; fieldtrip.fcdonders.nl). Trials containing horizontal eye movements were inspected manually and rejected if lateral eye fixations exceeded 1 sec during the Attention Task (3.55% of trials) or 150 msec during the final Awareness Task (12% of trials). Two faulty electrodes (across two participants) were interpolated using the nearest neighboring electrodes. Scalp electrode data were rereferenced to the average of all 64 electrodes, resampled to 256 Hz, and subjected to a surface Laplacian filter (Cohen, 2014). Trials were epoched into intervals containing signal at full coherence (Awareness Task: 1.4–3.4 sec or 4.3–6.3 sec, Figure 2; Attention Task: 1.9–11.9 sec, Figure 3) for frequency power analyses (see Results). Attention Task trials were also epoched with an additional 2 sec before and after each signal period for time–frequency power analyses.
Phase-locked Power Calculation
To measure neural responses to flickering stimuli in the Attention and Awareness Tasks, we examined phase-locked power (sometimes called “evoked” power) at each of the noise (10 and 15 Hz) and signal (5 and 7.5 Hz) stimulation frequencies. We elected to use phase-locked power as our measure of interest because it is maximally sensitive to neural responses in phase with the events of interest—in our case, the onsets of flickering images—and parcels out these responses from non-phase-locked neural activity (sometimes called “induced” power) that might otherwise obscure weak neural responses to invisible signals.
Phase-locked power was calculated as the difference between normalized total power and non-phase-locked power (Cohen, 2014). Total raw power was computed by applying Fourier transforms (Hanning window, 0.10-Hz frequency resolution) to 10-sec trial epochs in the Attention Task (1.9–11.9 sec; Figure 3) and 2-sec interval epochs in the Awareness Task (1.4–3.4 sec and 4.3–6.3 sec, Figure 2; zero-padded to 10 sec) and averaging across trials in each condition of interest (attention, awareness, stimulation frequency, and side). Total power in each condition was then decibel-normalized by dividing the raw power in each frequency bin by the average power in the 20 adjacent frequency bins (±1.0 Hz) and multiplying the logarithmic transform of the result by 10 (Cohen, 2014). Non-phase-locked power was calculated in the same manner as total power, after the condition average ERP had been subtracted from each trial (Cohen, 2014). Finally, phase-locked power (hereafter referred to as “power”) was calculated by subtracting the non-phase-locked power from the total power within each condition.
To test whether participants maintained covert attention during the Attention Task, we also calculated noise frequency power as a function of time. Preprocessed EEG data were bandpass filtered at each frequency of interest (MATLAB function: fir1, order: 64 samples, width = 0.01 Hz), subjected to a Hilbert transform, and down-sampled to 40 Hz. Phase-locked time–frequency power was then calculated in the same manner as phase-locked frequency power (above).
To maximize power for all statistical analyses, we subjected the data to a contralateralization procedure to remove the side of stimulation (left or right of fixation) as a factor within each attention and awareness condition. The electrode labels in trials with right-sided stimulation (i.e., when stimuli on the right of fixation flickered at the frequency of interest) were mirrored along the sagittal center line (e.g., PO7 became PO8, and vice versa). After this procedure, left-sided electrodes in all trials (irrespective of stimulation side) represented those ipsilateral to stimulation, and right-sided electrodes represented those contralateral to stimulation. Because hemispheric differences were not crucial to our research question, we then collapsed power across the factor of stimulation side. All electrode topographies presented here (Figures 6, 7, 8, and 10) represent data that underwent this contralateralization procedure.
The initial adaptive staircase procedure produced an average signal coherence of 29.91% (SD = 3.18%) for the invisible condition and 69.91% (SD = 3.18%) for the visible condition, across participants. One-tailed t tests were used to assess signal awareness in the final Awareness Task, which revealed that visible targets were detected above chance (chance = 50%; M = 95.77%, SD = 3.64%, t(22) = 60.367, p < .001) and that invisible targets were detected no better than chance (M = 50.96%, SD = 8.13%, t(22) = 0.565, p = .289). Furthermore, Bayesian statistics supported the null hypothesis that invisible stimuli were detected at chance (uniform prior, lower bound = 50%, upper bound = 100%, B = .07).
One-tailed t tests revealed that contrast decrement targets were detected better than chance level (chance = 33%; M = 65.72%, SD = 6.77%, t(22) = 46.302, p < .001). A two-tailed t test revealed that contrast decrement targets were better detected when the signal was visible (M = 68.11%, SD = 8.38%) than when it was invisible (M = 63.34%, SD = 5.71%, t(22) = 4.84, p < .001).
Noise and Signal Elicit Distinct Neural Responses
To confirm that our measure of phase-locked power successfully isolated neural responses to signal and noise stimuli, we computed power in the Attention Task (see Methods) and collapsed across awareness conditions and participants. Figure 5 shows the phase-locked power at contralateral electrode PO3/4 as a function of frequency, separately for each combination of stimulation frequencies. Note that power is only greater than zero at the signal (5 and 7.5 Hz) and noise (10 and 15 Hz) frequencies, confirming that the measure successfully isolated neural responses to the flickering stimuli.
Spatial Attention Enhances Neural Representations of Noise
To verify that covert attention was directed to the cued image stream (left or right) throughout Attention Task epochs, we assessed differences in time–frequency power between attended (cued) and ignored image streams. Time–frequency power was computed using Hilbert transforms (see Methods) and collapsed across noise frequencies and awareness conditions (because all stimuli contained noise). The effect of attention was then tested with a two-tailed Monte Carlo cluster-based permutation test in the Fieldtrip toolbox for MATLAB (between participant factors: electrode power and time, cluster p < .05, unit p < .05, 1000 permutations). Cluster-based permutation analyses are a nonparametric method for testing condition differences in high-dimensional neural data while correcting for multiple comparisons (for a detailed discussion, see Maris & Oostenveld, 2007). They are typically most useful when experimenters have few a priori expectations about specific locations or times of effects (Groppe, Urbach, & Kutas, 2011), as was the case in the current investigation. As revealed in Figure 6, spatial attention enhanced noise frequency power across a cluster of posterior and contralateral electrodes that spanned the entire epoch (cluster-corrected p = .002).
Target Detection Correlates with the Effect of Attention on Neural Representations of Noise
Next, we investigated the relationship between behavioral performance on the Attention Task and the effect of attention on neural representations of noise stimuli. We labeled trials in which participants identified the exact number of targets (0, 1, or 2) as correct and all other trials as incorrect and then balanced the number of correct and incorrect trials in each condition by removing a random subset of trials in the category with the greater number of trials. Noise frequency power was computed (see Methods) and collapsed across frequencies and sides, and the effect of attention was computed as the difference between attended and ignored trials (attended − ignored). Finally, the attentional modulation of correct and incorrect trials was compared with a two-tailed Monte Carlo cluster-based permutation test (between participant factor: electrode power, cluster p < .05, unit p < .05, 1000 permutations). As can be seen in Figure 7, there was a larger effect of attention on the neural response to noise stimuli across frontal and central electrodes when targets (contrast decrements) were correctly detected (cluster-corrected p = .014).
Invisible Signals Elicit Reliable Frequency Responses
A central goal of our study was to determine whether invisible (and visible) signals elicit reliable SSVEPs. To do this, we calculated power at the signal frequencies (5 and 7.5 Hz; see Methods) and collapsed across frequencies and attention conditions. We then compared the electrode distributions to a zero power electrode distribution with a one-tailed Monte Carlo cluster-based permutation test (between-participant factor: electrode power, cluster p < .05, unit p < .05, 1000 permutations), separately for each level of awareness. As revealed in Figure 8, signal frequency power during presentation of a visible signal was significantly greater than zero across a broad posterior and mostly contralateral cluster of electrodes (cluster-corrected p = .002), confirming the presence of a neural response to visible signals. Crucially, signal frequency power during presentation of invisible signals was also significantly greater than zero across a cluster of posterior and mostly contralateral electrodes (cluster-corrected p = .002), confirming the presence of a neural response to invisible signals.
Signal Frequency Responses Are Not Driven by Noise Stimuli
As a control, we checked whether the neural activity observed at signal frequencies might reflect a neural response to noise stimuli at half the frequency of stimulation. To do this, we computed frequency power in Awareness Task intervals (see Methods) and collapsed across the cluster of electrodes that showed a significant response to invisible stimuli in the Attention Task (Pz, POz, Oz, PO3, PO4, contralateral PO7/8, contralateral O1/2, ipsilateral P1/2, and ipsilateral P3/4; see Figure 8), separately for intervals that contained signal and those that contained only noise (at each frequency of interest). As can be seen in Figure 9, Awareness Task intervals that contained signal (gray lines) elicited peaks in the frequency spectrum at signal frequencies (5 or 7.5 Hz), but intervals that contained only noise (black lines) produced no such activity (signal interval > noise interval, paired samples t test, t = 5.593, p < .001). One-tailed t tests demonstrated that signal frequency power was greater than zero during signal intervals (M = 0.08 dB, SD = 0.07 dB, t = 5.931, p < .001), but no greater than zero during noise-only intervals (M < 0.01 dB, SD = 0.01 dB, t = 0.965, p = .172). Crucially, Bayesian statistics supported the null hypothesis that noise stimuli produced no neural response at signal frequencies (uniform prior, lower bound = 0, upper bound = 0.08 dB, B = .10). This finding aligns with that of a previous study that used an analogous frequency tagging approach (Ales et al., 2012), which demonstrated no neural response at the signal frequency when participants were shown image sequences with noise images only. Together, these results confirm that the observed neural activity at signal frequencies in the Attention Task was driven by signal stimuli and not a subharmonic response to noise stimuli.
Attention Enhances Neural Representations of Visible and Invisible Signals
Considering the weaker neural response to signals relative to high-contrast noise (Figure 5), we collapsed power across posterior and contralateral clusters of electrodes that showed a significant response to the signal (Figure 8), separately for each level of awareness and attention. As revealed in Figure 10, attention increased the neural response to both visible and invisible signals across these electrode clusters. A two-way ANOVA tested the effects of signal coherence (two levels: visible, invisible) and spatial attention (two levels: attended, ignored) on neural responses to signal. Results of the ANOVA revealed a main effect of signal coherence, F(1, 22) = 47.699, p < .001, ηp2 = .68, with greater neural responses to visible signals (M = 0.35 dB, SD = 0.25 dB) than to invisible signals (M = 0.05 dB, SD = 0.05 dB). Spatial attention also increased neural responses to stimuli, F(1, 22) = 7.693, p = .011, ηp2 = .26, with significantly greater signal frequency power in response to attended signals (M = 0.24 dB, SD = 0.16 dB) than ignored signals (M = 0.16 dB, SD = 0.16 dB). The interaction between signal coherence and spatial attention was also significant, F(1, 22) = 4.768, p = .040, ηp2 = .18. Follow-up paired samples t tests revealed that the interaction was driven by a greater effect of attention on visible signals (ΔM = 0.11 dB, ΔSD = 0.19 dB) than invisible signals (ΔM = 0.04 dB, ΔSD = 0.08 dB, t(22) = 2.18, p = .040).
Because our critical research question related to whether attention can modulate neural responses to invisible stimuli, we also followed up the main effect of attention with t tests of the simple main effect of spatial attention at each level of signal awareness (Figure 10B). Spatial attention enhanced neural responses to visible signals, with greater activity in response to attended (M = 0.40 dB, SD = 0.26 dB) than ignored visible stimuli (M = 0.30 dB, SD = 0.27 dB, t(22) = 2.671, p = .014). This finding is in line with previous research showing attentional enhancement of SSVEPs to visible flickering stimuli (Vialatte, Maurice, Dauwels, & Cichocki, 2010). Crucially, spatial attention also modulated neural responses to invisible signals, with significantly greater activity in response to attended (M = 0.07 dB, SD = 0.07 dB) than ignored invisible stimuli (M = 0.03 dB, SD = 0.06 dB, t(22) = 2.363, p = .027), indicating that attention can also enhance neural representations of invisible stimuli embedded in highly salient noise.
Previous research has suggested that covert spatial attention can modulate neural processing of invisible stimuli, supporting the notion that attention and awareness are dissociable neural mechanisms (Wyart et al., 2012; Watanabe et al., 2011; Wyart & Tallon-Baudry, 2008). Nevertheless, the intricacies of such a relationship remain poorly understood. In particular, no study to date has demonstrated that spatial attention can modulate neural representations of invisible stimuli or assessed the nature of such modulation when those stimuli are in spatial competition with highly salient noise. To investigate these questions, we developed a novel attention task in which participants counted the number of brief contrast decreases in one of two image streams that contained both signals (visible or invisible) and noise. We isolated neural responses to noise in cued (attended) and noncued (ignored) image streams and observed enhanced activity across contralateral and posterior electrodes to cued noise throughout the trial epoch, confirming that participants voluntarily held their attention to one of the two lateralized image streams as instructed. The effect of attention on neural responses to noise was also greater across frontal and central electrodes with correct identification of contrast targets (Figure 7), suggesting that fluctuations in attention across trials directly affected target detection.
We employed a novel frequency tagging approach that allowed us to isolate neural representations of visible and invisible signals embedded in highly salient noise. To our knowledge, this is the first study to report SSVEP responses to objectively invisible stimuli embedded in noise. It could be argued that, because we did not measure signal awareness during the Attention Task, participants might have been aware of the “invisible” signal. Although we cannot rule this out, such a scenario is highly unlikely, considering that participants actively searched for signals in the Awareness Task but looked instead for contrast decrements during the Attention Task. Thus, our results suggest that awareness of a masked stimulus is not a prerequisite for eliciting an SSVEP, as might be inferred from the step-like rise in SSVEP power that coincided with the onset of signal awareness in a previous study (Ales et al., 2012). Instead, our findings demonstrate that SSVEPs track intermediate levels of signal strength, even at levels too weak to provoke conscious perception.
Critically, our paradigm allowed us to measure the effects of spatial attention on neural representations of visible and invisible signals. We found that neural representations of visible signals were greater in the attended image stream than in the ignored stream, extending previous findings of effects of attention on neural representations of visible stimuli (Martinez et al., 1999; Hillyard & Anllo-Vento, 1998; Müller et al., 1998) to demonstrate that spatial attention also benefits partially degraded, yet still visible, signals in spatial competition with clearly visible and highly salient noise. Crucially, neural responses to invisible signals were also greater in the attended image stream than in the ignored stream, demonstrating that spatial attention enhances representations of invisible stimuli in direct spatial competition with highly salient and visible noise. Because spatial attention enhanced neural representations of signals without a corresponding increase in signal awareness, the present findings support the notion that spatial attention and awareness are dissociable neural mechanisms (Cohen, Cavanagh, Chun, & Nakayama, 2012; Koch & Tsuchiya, 2012; Tallon-Baudry, 2012; Dehaene, Changeux, Naccache, Sackur, & Sergent, 2006).
Although this study is not the first to demonstrate effects of spatial attention in the absence of object awareness (Wyart et al., 2012; Watanabe et al., 2011; Schurger et al., 2008; Wyart & Tallon-Baudry, 2008), it makes several important advances on the existing literature. First, this study investigated a distinct question to that of previous studies that aimed to assess the effects of both awareness and attention on neural processing, independent of signal strength (Wyart et al., 2012; Schurger et al., 2008; Wyart & Tallon-Baudry, 2008). In these previous studies, signals were presented at detection threshold, and participants' subjective reports were used to categorize trials as visible or invisible. The authors found effects of attention on the neural processing of perithreshold signals, even when participants reported being unaware of their presence. As such, these studies provide evidence that attention can modulate perithreshold stimuli but cannot speak to how the visual system treats very weak signals with insufficient bottom–up activation to enter awareness, irrespective of the cognitive state of the observer (so-called “subliminal” stimuli; Dehaene et al., 2006). In this study, we presented visible and invisible signals at different, predetermined levels of coherence and verified that invisible stimuli were objectively undetectable with a two-interval forced-choice signal detection task. Thus, we can be confident that the invisible stimuli in our experiment were not perceived due to a lack of bottom–up activation, rather than fluctuations in the cognitive state of the observer. Correspondingly, our findings demonstrate that neural processing of objectively subliminal stimuli can be modulated by spatial attention, as suggested by Dehaene and colleagues (2006), and that surpassing a hypothetical “threshold” is not a necessary precursor for modulation by spatial attention.
Second, previous studies have not demonstrated that the observed neural activity, modulated by attention, was specifically related to the invisible stimuli in question. As such, previously observed effects of attention may instead reflect (a) baseline shifts in neuronal activity that occur even in the absence of external driving stimuli (as may be the case in Watanabe et al., 2011; see Driver & Frith, 2000); (b) enhanced neural representations of other visible stimuli (e.g., the spatial cue in Wyart et al., 2012, as has been argued by Cohen et al., 2012); or (c) subcomponents of spatial attention that do not modulate neural representations per se (e.g., spatial reorienting after a miscued stimulus in Wyart et al., 2012; Schurger et al., 2008; Wyart & Tallon-Baudry, 2008). In demonstrating that spatial attention modulates specific neural correlates of invisible stimuli, without a corresponding increase in awareness, this study provides the first clear evidence that spatial attention and awareness dissociate at the level of neuronal representations.
A third and, arguably, most important advance of the current study is that we have shown here that spatial attention can enhance neural representations of invisible stimuli that are in direct spatial competition with highly salient, visible stimuli. Previous studies presented invisible signals alone (Schurger et al., 2008; Wyart & Tallon-Baudry, 2008) or at different times or locations (Wyart et al., 2012; Watanabe et al., 2011) to the salient masks used to titrate signal awareness. Because neural competition is maximal at the level of the receptive field (Beck & Kastner, 2009; Reynolds, Chelazzi, & Desimone, 1999), neural representations of invisible signals in these studies were likely under conditions of minimal competition. In contrast, we maximized competition between signal and noise by presenting them concurrently and at the same location. Our findings reveal concurrent neural representations of both visible and invisible stimuli at the same location, demonstrating that spatial competition with highly salient stimuli is not sufficient to suppress weak neural representations of invisible stimuli. Moreover, this study demonstrates that weak neural representations of invisible stimuli in competition with salient stimuli can nevertheless be biased according to the top–down goals of the observer—in this case, holding covert attention preferentially to the left or right visual field. Given that signal features were irrelevant to the contrast detection task, this finding suggests that all stimuli at attended locations are prioritized relative to those at unattended locations, irrespective of their task relevance, their capacity to enter awareness, or their proximity to more salient stimuli.
The present findings demonstrate that spatial attention can operate independent of mechanisms of awareness at the level of neural representations. More broadly, the present findings place spatial attention within a growing body of literature that suggests various forms of attention (e.g., temporal, feature-based, and involuntary spatial attention) can operate in the absence of stimulus awareness (for a review, see Koch & Tsuchiya, 2007). Together, these findings argue against the idea that attention and awareness are identical (Prinz, 2012) and instead support theories that cast attention and awareness as dissociable mechanisms (Cohen et al., 2012; Koch & Tsuchiya, 2012; Tallon-Baudry, 2012; Dehaene et al., 2006). Nevertheless, the exact nature of this relationship remains to be fully characterized, in particular whether the different forms of attention interact with awareness according to the same underlying principles and how such top–down biases interact with bottom–up processes related to salience and neural competition between representations. To this end, we anticipate that the present paradigm could be adapted to study how other nonspatial forms of attention (e.g., feature-based attention) modulate neural representations of multiple competing stimuli at different levels of awareness.
This research was supported by the Australian Research Council (ARC) Centre of Excellence for Integrative Brain Function (ARC Centre Grant CE140100007). J. B. M. was supported by an ARC Australian Laureate Fellowship (FL110100103).
Reprint requests should be sent to Cooper A. Smout, QBI Building 79, University of Queensland, St. Lucia QLD 4072, Australia, or via e-mail: firstname.lastname@example.org.