The human visual system differentially represents subjectively and objectively invisible stimuli

The study of unconscious processing requires a measure of conscious awareness. Awareness measures can be either subjective (based on participant’s report) or objective (based on perceptual performance). The preferred awareness measure depends on the theoretical position about consciousness and may influence conclusions about the extent of unconscious processing and about the neural correlates of consciousness. We obtained functional magnetic resonance imaging (fMRI) measurements from 43 subjects while they viewed masked faces and houses that were either subjectively or objectively invisible. Even for objectively invisible (perceptually indiscriminable) stimuli, we found significant category information in both early, lower-level visual areas and in higher-level visual cortex, although representations in anterior, category-selective ventrotemporal areas were less robust. For subjectively invisible stimuli, similar to visible stimuli, there was a clear posterior-to-anterior gradient in visual cortex, with stronger category information in ventrotemporal cortex than in early visual cortex. For objectively invisible stimuli, however, category information remained virtually unchanged from early visual cortex to object- and category-selective visual areas. These results demonstrate that although both objectively and subjectively invisible stimuli are represented in visual cortex, the extent of unconscious information processing is influenced by the measurement approach. Furthermore, our data show that subjective and objective approaches are associated with different neural correlates of consciousness and thus have implications for neural theories of consciousness.

In the masking-efficiency experiment, at the three highest mask strengths, where average d' was below one, between-subject variability was low (SDs for mask contrast of 38, 61, and 100% were 0.41, 0.31, and 0.41, respectively, see S1 Fig): different observers performed in a narrow range for high mask contrast. We therefore calibrated mask contrast for the whole group of participants in the fMRI experiment based on the masking-efficiency experiment group data. An alternative approach is to calibrate mask contrast individually to allow for maximum signal strength for every participant. Such individual calibration requires reliable individual-subject estimates. However, the data from the masking-efficiency experiment revealed low reliability for individual differences at high mask contrasts. S1 Table shows correlations of face-house discrimination performance (in proportion correct, to avoid extreme d' values for small number of trials) between the nine different mask contrasts. Note that for the three mask levels where d' was below one correlations are below r = .17. The diagonal of S1 Table   shows within-condition reliability estimates. These were calculated by correlating performance from two randomly determined halves of the data set. This process was repeated 100 times; the resulting correlation coefficients were averaged (after Fisher z-transformation), and the mean score was backtransformed. While reliability was generally high when overall performance was above a d' of one, for the three highest masking strength reliability was low (below r = .19). With such low reliability for individual differences participant-based adjustment of mask contrast (for example, via an adaptive staircase procedure) would capitalize on noise. This, in turn, would have risked above-chance performance for some participants in the fMRI experiment. To avoid issues with post-hoc data selection 1 , such as regression to the mean 2 , we considered group-based calibration of mask contrast superior for the purpose of our study. Note that these considerations are naturally limited to the particular stimuli and sandwich masking paradigm adopted in our study; with other presentation methods, such as continuous flash suppression 3 , individual stimulus calibration may be preferable.

Evidence for objective invisibility
For the interpretation of our fMRI results in our obj-inv condition it is important to test whether faces and houses were truly indiscriminable (objectively invisible). In the main manuscript, we report that discrimination performance (M = 0.02, SD = 0.24) did not differ significantly from chance, with moderate evidence for the null hypothesis of chance-level discrimination (t(42) = 0.45, p = .33 (onetailed), dz = 0.07, BF0+ = 4.13). Standard statistical significance testing. Another approach that has been adopted in the literature on unconscious processing is to test every participant's performance against chance level. S2 Fig shows face-house discrimination as proportion correct. One-tailed binomial tests comparing every participant's accuracy against the chance level of 0.5 revealed (only) one participant whose performance was significantly above the chance level (p = .028). When carrying out 43 tests, some are expected to return a significant result. In fact, the distribution of p-values from the 43 binomial tests (S1 Fig) looks just as one would expect if values were randomly sampled from a binomial distribution with a probability of 0.5 success for each trial. However, as discussed in the main paper, although absence of awareness is commonly demonstrated by showing that performance did not exceed what is expected by chance (p > .05), a non-significant result simply means that the observed effect is not very surprising under the null hypothesis of no effect, but cannot be taken as support for the null hypothesis.
Furthermore, p-values increase with decreasing power and increasing variance in the awareness measure, such that a non-significant result is not particularly surprising if the awareness measure is collected in few trials and small participant samples (i.e. with low power) 4,5 . To address these issues, we tested a much larger sample than previous fMRI studies, measured awareness in every trial during scanning (average of 196 trials per participant to calculate d' in the obj-inv condition), and did not exclude any participants based on their awareness data (which can severely inflate estimates of unconscious processing due to regression to the mean 2 ). Although our fMRI study adopted the highest criteria for demonstrating unconscious processing of objectively invisible stimuli to date, approaches other than the standard frequentist significance-testing approach are required to provide evidence for absence of awareness. Two such alternative approaches are Bayesian statistics (as reported in the main paper) and equivalence tests.
Equivalence tests. Equivalence tests based on the two one-sided tests (TOST) procedure can be used to determine if an observed effect is surprisingly small under the assumption that a true effect, at least as large as a specific smallest effect size of interest (SESOI), exists 6 . In the context of unconscious processing, setting an SESOI represents a non-trivial, subjective call, as it requires determining the smallest d' score that can still reasonably be considered objectively invisible. Setting the SESOI (arbitrarily) to a d' score of 0.10 the TOST procedure (using the spreadsheet TOST calculators provided by Lakens and colleagues 6 ) revealed that the observed effect was significant within the equivalent bounds of -0.1 and 0.1 scale points, t(42) = -2.23, p = .016. We note that when setting the SESOI to a d' score of 0.15, the TOST would return p < .001; when setting the SESOI to a d' score of 0.05, however, the TOST would return p = .19. These results provide more insight about the uncertainty in our obj-inv data. While we cannot reasonably claim that discrimination performance was surprisingly low if the true effect was Bayesian statistics. The Bayesian statistical framework allows estimating how likely the observed data are under the null hypothesis, compared to the alternative hypothesis, and has thus particular utility in the study of unconscious processes 7 . While the null hypothesis in this framework is a point estimate (zero), Bayesian t-tests also require a non-trivial, subjective call in specifying the prior distribution for the alternative hypothesis. Estimates of the probability of the null hypothesis relative to the alternative hypothesis (BF0+) depend on this prior. In the main paper, we used the default prior provided in the JASP software package (Cauchy distribution centered on zero and scale 0.707), and this resulted in "moderate" evidence for the null hypothesis (BF0+ = 4.13). A wider prior (assigning more weight to larger effect sizes) would result in higher probability of the null hypothesis (e.g. with a Cauchy distribution with a scale of one, BF0+ = 5.67). Conversely, a narrower prior (assigning more weight to smaller effect sizes) would results in lower probability of the null hypothesis (e.g. with a normal distribution with M = 0 and SD = 0.5, BF0+ = 2.34). Rather than allowing definite conclusions about absence of awareness, our results highlight the statistical uncertainty inherent in demonstrations of unconscious processing, and invite discussions about reasonable prior assumptions and about which (bands of) effect sizes represent conclusive evidence for presence vs. absence of conscious awareness.

Univariate ROI results
We also conducted standard univariate ROI analyses (same ROI definitions as for the MVPA analyses).
Here, for every participant and ROI we used the localizer data to determine the 100 most faceresponsive voxels and the 100 most house-selective voxels. We then calculated for every visibility condition in the main experiment the mean difference in activity evoked by faces vs. houses in the faceresponsive voxels and the mean difference in activity evoked by houses vs. faces in the house-responsive voxels. These differences were then averaged to yield a mean beta difference (Δb) reflecting how well a region's activity in the main experiment distinguished its preferred category from the non-preferred category; Δb was used as a univariate measure of category information. Univariate results were similar to the MVPA results reported in the main paper (compare S3 Fig with Fig 2).  62´10 3 , S3 Fig). For obj-inv, this gradient was virtually absent; category information did not differ significantly between V1 and LOC (t(42) = 0.06, p = .95, dz < 0.01, BF01 = 6.05) or between OFA/OPA and FFA/PPA (t(42) = 0.07, p = .95, dz = 0.01, BF01 = 6.04).

Different ROI definitions
For the univariate analyses and for the ROI analyses reported in the main paper we selected the 100 most face-responsive and the 100 most house-responsive voxels for each ROI, as determined by the localizer scan. To ensure that our results did not depend on this particular ROI size of 200 voxels, we repeated the MVPA correlation analyses for other ROI definitions (including between 10 [i.e. the 5 most face-responsive, and the 5 most house-responsive voxels in the localizer] and 500 voxels, in steps of 10).
As can be seen in S4 Fig, in all ROIs except for V1, overall pattern discrimination (category information Δr) dramatically increased with smaller ROI size. This was especially the case for the obj-vis and subjvis conditions, and in LOC and in FFA/PPA also for subj-inv. ROI size had little influence on category information for the obj-inv condition. Most importantly, the patterns of differences between visibility conditions as reported in the main paper were robust across ROI definitions.

Main-experiment searchlights
All analyses reported in the main paper relate brain activity in the main experiment to the independent localizer scan, where participants did a simple one-back task on unmasked faces and houses presented for 750 ms each in a block design. We designed the study for this particular analysis strategy, because it is statistically powerful and isolates perceptual representations from later cognitive processes related to the task, decision, and motor processes 8,9 . Analyses based on the data from the main experiment alone cannot exclude such influences, for example, in these analyses button presses are confounded with the different stimulus and visibility conditions. Category information. Accordingly, additional searchlight analyses based on the mainexperiment data alone revealed category information in motor cortex, driven by the responses given to the presented categories (S5 Fig). For these searchlights, the main experiment data were split into two sets (all possible splits of two and three runs). For every visibility conditions we repeatedly calculated within-and between-category pattern correlation differences for a moving sphere with a radius of five voxels (524 voxels in the sphere) which was centered on every voxel in the functional images of every participant. Correlations from all possible splits were then averaged to yield the searchlight maps.
These analyses revealed areas carrying significant category information across the whole brain (corrected for multiple comparisons via false discovery estimation, p < .05) in the obj-vis, subj-vis, and subj-inv condition, but not in the obj-inv condition (S5 Fig). In all three visibility conditions clusters with significant category information were located in bilateral fusiform gyrus, but most prominently in bilateral motor cortex (highlighted in the slices shown in S5 Fig). Such motor cortex pattern discrimination reflects the correlation between stimulus category and participant's button presses for those conditions in which button presses (perceptual discrimination) carried information about stimulus category (when discrimination performance is above chance). Accordingly, additional frontal cortex clusters (in right inferior frontal gyrus, orbital parts of right superior and left middle frontal gyrus) revealed in the contrast between subj-inv and subj-vis (S5 Fig) are difficult to interpret. Remarkably, there was no evidence for category information in motor cortex in the obj-inv condition, indicating that in this condition (neural representations of) button presses were indeed unrelated to stimulus category, thus further supporting the notion that stimuli were fully indiscriminable.
Visibility irrespective of category. In addition to occipitotemporal areas, several previous fMRI studies reported correlates of conscious visual perception in a frontoparietal network 10,11 .These studies compared brain activity evoked by visible and invisible conditions, independent of stimulus category or content, assessing visibility either with objective measures (e.g. by comparing weakly vs. strongly masked stimuli 12 ) or with subjective measures of visibility (e.g. by comparing similar or identical stimuli judged as "visible" vs. "invisible" 11 ). To test for such frontoparietal involvement in content-independent conscious perception ("conscious access" or "report"), we conducted searchlight analyses that revealed brain areas where activity distinguished between visibility conditions, independent of object category.
For these searchlights, we again split the main experiment data into two sets (all possible splits of two and three runs). Separately for the objective and for the subjective condition we repeatedly calculated within-visibility (visible/visible and invisible/invisible) and between-visibility (visible/invisible) pattern correlations for a moving sphere with a radius of five voxels (524 voxels in the sphere) which was centered on every voxel in the functional images of every participant. Correlations from all possible splits were then averaged to yield the searchlight maps.
In both the objective and subjective condition activity in many brain areas distinguished between visible and invisible trials (corrected for multiple comparisons via false discovery estimation, p < .05), including early visual cortex, fusiform and inferior temporal gyrus, parietal cortex (e.g. inferior parietal lobule, precuneus, angular gyrus, postcentral gyrus), frontal cortex (e.g. motor cortex, inferior frontal gyrus, superior frontal gyrus, middle frontal gyrus, anterior cingulate), and anterior cingulate cortex. The sagittal views of the right hemisphere in S6 Fig highlight significant pattern correlations in ventrotemporal areas, parietal cortex, and inferior frontal gyrus in both objective and subjective conditions. These results could be considered evidence for frontoparietal involvement in (contentindependent) conscious perception. Note, however, that in these analyses based on the main-experiment data, post-perceptual processes (related to the task, decision, and motor processes) and stimulus strength (mask contrast) are confounded with differences in visibility