Our senses constantly receive information from the world around us, but we consciously perceive only a small portion of it. Nonetheless, even stimuli that are not consciously perceived are registered in our brain and influence our behavior. This is known as unconscious perception.

Researchers disagree about how brain activity differs during conscious and unconscious perception. Some think that both consciously and unconsciously perceived objects are processed in the same way in the brain, but that the brain is more active during conscious perception. Others think that different neurons process the information in different types of perception.

Salti et al. have now investigated this issue. While recording participants' brain activity, a line was briefly presented in one of eight different possible locations on a screen. The line was masked so it would be consciously perceived in roughly half of the presentations. Participants had to report the location of the line and then say whether they had seen it or had merely guessed its location. Even when they reported that they were guessing, participants identified the location of the line better than by chance, indicating unconscious perception on ‘guess’ trials. This enabled Salti et al. to compare how the brain encodes consciously perceived and unconsciously perceived stimuli.

Unlike previous studies in which the brain activity associated with ‘seen’ and ‘unseen’ stimuli was compared, Salti et al. used a different approach to extract the neural activity underlying consciousness. A classifying algorithm was trained on a subset of the data to recognize from the recorded brain activity where on the screen a line had appeared. Applying this algorithm to the remaining data revealed the dynamics of stimulus encoding. Consciously and unconsciously perceived stimuli are encoded by the same neural responses for about a quater of a second. From this point on, consciously perceived stimuli benefit from a series of additional brain processes, each restricted in time. For unconsciously perceived stimuli, this chain of processing breaks and a slow decay of encoding is observed.

Salti et al., therefore, conclude that conscious perception is represented differently to unconscious perception in the brain and produces more extensive and structured brain activity. Future work will focus on understanding these differences in neural coding and their contribution to the interplay between conscious and unconscious perception.

The scientific investigation of consciousness divides into two major branches (Koch, 2004; Dehaene et al., 2014). The first branch, which studies the ‘state of consciousness’, focuses on general vigilance or a person's ability to perceive, interact, and communicate with the environment, by examining the regulation of sleep and waking, and their pathological disruption by coma, epilepsy, and sleep disorders (Laureys et al., 2004). The second branch of inquiry, which studies ‘access to consciousness’, focuses on the processes that make a specific content subjectively experienced—what differences in brain activity distinguish conscious vs unattended or subliminal stimuli (Dehaene et al., 2006). To a large extent, both endeavors rely on characterization of the neural activity that underlies the conscious state or conscious perception, and they are traditionally approached in a similar manner (Dehaene and Changeux, 2011). In order to examine the neuronal processes that underlie conscious states, the neuronal activity associated with one state of consciousness (e.g., sleep) is compared to another (e.g., awake). Similarly, in order to extract the neural correlates of conscious perception, threshold-level stimuli are presented to the subject such that they are sometimes consciously perceived and sometimes not. Trials are then sorted according to their subjective perception (‘Seen’ or ‘Unseen’), and the neurophysiological signatures associated with these categories are contrasted to extract the correlates of consciousness perception.

Such paradigms have yielded convergent findings (Koch, 2004; Dehaene and Changeux, 2011; Dehaene et al., 2014), but they have also been criticized on several grounds (Aru et al., 2012). First, differences in neural activity may stem from physical differences in the stimuli. To control this, many experiments now employ identical stimuli and rely on participants' subjective reports to differentiate seen and unseen conditions. Even then, stimuli also differ in their depth of processing: performance is typically much higher for conscious trials (Lau and Passingham, 2006), and several operations may only be feasible in the conscious state (Sackur and Dehaene, 2009; De Lange et al., 2011). The main effect of seen vs unseen trials may, therefore, reveal processing differences unrelated to the actual cerebral encoding of a conscious or unconscious stimulus, and reflecting solely, the operations that precede follow or even coincide with conscious access (Aru et al., 2012; Pitts et al., 2012; Frässle et al., 2014; Pitts et al., 2014).

Another limitation is that these paradigms do not isolate the brain mechanisms underlying the conscious representation of a specific content (Haynes, 2009). The broad contrast between brain activity on ‘seen’ vs ‘unseen’ trials may include generic processes of attention, alerting, or reporting that should be carefully kept distinct from the more limited set of processes that represent the current mental content. For instance, the late P3 component of event-related potentials, which is a frequent signature of conscious perception (Dehaene and Changeux, 2011), has been proposed to reflect either a non-specific alerting effect arising from noradrenergic neurons of the locus coeruleus (Nieuwenhuis et al., 2005) or the activation of generic reporting processes unnecessary to conscious perception itself (Pitts et al., 2012, 2014). To resolve this issue, it is necessary to identify which patterns of brain activity, unique to conscious trials, faithfully encode the contents of subjective experience and to separate them from other non-specific brain responses (Haynes, 2009).

The presence or absence of additional information is a crucial feature that separates major theories of conscious access. Some theories propose that subjective experience emerges from a generic pattern of brain activity without any stimulus-specific change (Lau and Rosenthal, 2011). According to these models, the cognitive representation of a stimulus is built during unconscious perception, and the additional activity that gives rise to consciousness tags it as ‘perceived’ but adds no perceptual or coding benefit. Conversely, other models (Dehaene et al., 2003, 2006) suggest that conscious perception relies on a massive amplification and cortical broadcasting of stimulus-specific information. The Global Neuronal Workspace (GNW) model asserts that perceptual stages may unfold identically on conscious and non-conscious trials, but that, when a stimulus gains access to consciousness, stimulus-specific information is amplified and re-encoded in additional areas, including a prefrontal–parietal network. This ‘workspace’ maintains the information online and dispatches it to additional processors, thus making the stimulus reportable. This view predicts that, ceteris paribus, additional stimulus information should be present on conscious trials. Finally, theories that associate conscious perception with posterior visual loops (Lamme, 2006) or with the integration of information into a coherent whole (Oizumi et al., 2014) are agnostic with respect to this issue: although greater reverberation or integration may cause an amplification of neuronal activity, (Fahrenfort et al., 2012) it could also, on the contrary, lead to diminished or sparser activity due to the greater predictability of the distributed brain signals being integrated (Friston, 2005).

In summary, for both empirical and theoretical reasons, it is essential to study the temporal propagation of conscious vs unconscious information in the brain and to probe the presence of additional stimulus-specific cortical codes on seen trials relative to unseen trials, while excluding any difference in perception and behavior. To fulfill this demanding agenda, we used time-resolved MEG and EEG recordings to track the internal representation of a flashed stimulus under visible and invisible conditions. This strategy was made possible by advances in multivariate decoding, which can detect stimulus-specific information in brain activity, and evaluate its localization (Kamitani and Tong, 2005; Haynes and Rees, 2006; Haynes, 2009) and its time course (King and Dehaene, 2014). We chose visual location as the decoded parameter of the stimulus, because multiple macroscopic retinotopic maps are present throughout the human brain, including the frontal lobe (Sereno et al., 1995; Hagler and Sereno, 2006), and therefore, the full cortical processing stream of this parameter should be more easily decodable than other more microscopic parameters, such as object identity. To further address the concern of differences in processing depth, we placed ourselves under ‘blindsight’ conditions by asking subjects to perform a forced-choice location task in which performance was excellent even on subjectively invisible trials. Selecting only the correct trials allowed us to compare visible and invisible trials with identical stimuli and responses (Lamy et al., 2009).

We recorded Magnetoencephalography (MEG) and Electroencephalography (EEG) while presenting the subjects with a tilted line segment at one of eight possible locations arranged circularly around the fixation point. The target was followed by a mask whose contrast was adjusted individually for each participant to ensure conscious perception on roughly half of the trials. In each trial, participants provided three separate behavioral responses: 1) an immediate forced-choice localization of the target, 2) whether their first response was correct or not, and 3) whether they saw the target or not (see Figure 1A and ‘Materials and methods’). Participants were instructed to report the target as unseen only if they had no perception at all regarding its location. If they did not see the target they were asked to guess its location (response 1).

Figure 1

Download asset Open asset

Behaviorally, subjects reported seeing 62.4 ± 3.62% (mean ± sem) of presented targets, and the accuracy (i.e., reporting the correct location) in these trials was 86.6 ± 2.54%. Performance in the unseen trials was 50.7 ± 6.4% correct, the way higher than the chance level of 12.5% (t(11) = 5.73, p <0 0.001), indicating blindsight. At all eight stimulated locations, the dominant response was to correctly point to that location, even on unseen trials (see Figure 1B). Those ‘unseen–correct’ trials thus provide a minimal contrast with ‘seen–correct’ trials. Mean localization Respone Times (RTs) were 812.2 ± 44 ms. Localization RTs for different critical conditions were as follow: 701.4 ± 35 ms for ‘Seen–Correct’ trials, 878.1 ± 78 ms for ‘Unseen–Correct’ trials, and 1110 ± 101 ms for ‘Unseen-incorrect’ trials.

Time course of location decoding

To track the cortical representation of the target location through time, we trained multivariate pattern classifiers separately at every time sample to predict the target location from the recorded brain signals (see ‘Materials and methods’). For each time sample, we extracted classifiers' calibrated posterior probabilities (Platt, 1999) associated with the eight spatial locations. We first trained the classifiers on all trials, regardless of participants' responses. Figure 2A depicts the classifiers' assigned probability for the correct target location as a function of time. Four stages could be distinguished (this division will be used throughout our analyses). In a first stage (0–115 ms post stimulus), classification probability was at chance (0.1251 ± 0.0001; mean ± sem), t < 1. From 115 to 162 ms, decoding probability rose above chance and peaked sharply at 147 ms post stimuli (0.141 ± 0.0028), t(11) = 5.9, p < 0.0001. Then came a transition period (162–271 ms) with reduced but still significant decoding (0.137 ± 0.0023), t(11) = 5.48, p < 0.0001. Finally, in a fourth stage (271–800 ms), decoding probability rose again and remained above chance for about 500 ms (0.157 ± 0.0083), t(11) = 3.8, p < 0.0001.

Figure 2 with 4 supplements see all

Download asset Open asset

Asymmetrical generalization of decoding

Our basic analysis indicated that consciously perceived stimuli are coded more reliably than unconscious ones, but the nature of this difference could not be completely appreciated from this analysis. ‘Seen–Correct’ trials were more numerous than ‘Seen-incorrect’ trials, and this factor alone could perhaps explain why they were more efficiently decoded. In this case, if conscious and unconscious codes differ, then training specifically on unconscious trials should revert the pattern and yield better decoding on unseen compared to seen trials. The GNW model (Dehaene et al., 2003), however, makes a different prediction concerning the asymmetry of decoding on seen and unseen trials. Specifically, the model predicts that all the processing stages present on unseen trials should continue to be observed on seen trials, while the converse should not be true: on seen trials, there should be brain activity encoding the perceived stimulus within the GNW and not present on unseen trials.

We examined these two options by testing for cross-condition generalization: we trained the decoder on a subset of trials (either the unseen–correct trials, or the seen–correct trials) and then tested for generalization, either to left-out trials within the same category or to trials belonging to the other category. One technical difficulty was that, since the experimental conditions were defined by subjects' responses, the number of trials at each location was no longer balanced within each of these conditions, and this imbalance affected the pre-stimulus bias of the 8-category classifier as well as its capacity to generalize. To address this problem, we turned to a simpler binary classifier, which was trained to simply sort the stimuli into left-hemifield vs right-hemifield stimuli. Within each subject, we matched the number of trials in these two classes and also selected equal numbers of seen–correct and unseen–correct trials. Figure 3 shows that this binary decoder, now operating with a chance level of 50%, still performed above chance with approximately the same time course as the 8-location decoder.

Figure 3

Download asset Open asset

The end result demonstrated the predicted asymmetry. When trained on ‘Unseen–Correct’ trials, classification generalized identically to ‘Seen–Correct’ trials, but when trained on ‘Seen–Correct’ trials, classification performance dropped when tested on ‘Unseen–Correct’ (Figure 3). This effect was apparent in two time windows, the third time window (178–225 ms; average classification probability for ‘Seen–Correct’ = 0.526 ± 0.0064; for ‘Unseen–Correct’ = 0.515 ± 0.005; t(11) = 2.8, p = 0.017), and the fourth late time window (272–800 ms; ‘Seen–Correct’ = 0.575 ± 0.017; ‘Unseen–Correct’ = 0.547 ± 0.016; t(11) = 2.44, p = 0.032). These analyses indicate that seen–correct trials contained the same decodable stimulus information as unseen–correct trials, plus additional information unique to conscious trials.

Decoding cortical sources

The GNW model predicts that the additional information associated with conscious representation should be jointly encoded in a network of distributed prefrontal and parietal regions. In order to assess the contribution of these brain areas to the encoding of consciously perceived contents, we modeled the brain activity in 68 regions of interest covering the whole cortex (See ‘Materials and methods’) and re-trained our decoders using just the distributed source signals from these regions. The results with eight output classes (Figure 4—figure supplement 1) confirmed the existence of two main stages: an early one (∼100–200 ms) where location information was essentially confined to occipital, ventral visual, and lateral and mesial parietal regions, and a later one (>250 ms) where it became highly distributed to multiple cortical areas, particularly in prefrontal and anterior cingulate cortex (Figure 4—figure supplement 1). To avoid an unnecessary increase in the number of statistical tests, we confined the subsequent analyses of seen and unseen trials, to four a priori regions of interest covering all areas of the dorsal visual pathway containing retinotopic maps for location (Sereno et al., 1995; Hagler and Sereno, 2006): pericalcarine, superior parietal cortex, rostral medial frontal cortex, and superior-lateral frontal cortex. We then repeated the generalization across conditions and procedures. This region-based decoding revealed the activation of successive spatial codes (Figure 4): both pericalcarine and superior parietal cortices contained decodable information as early as 115 ms post stimuli ([0.519 ± 0.005], t(11) = 3.82 p = 0.0028 and [0.516 ± 0.003], t(11) = 4.52 p < 0.0001, respectively). The two frontal areas revealed a more delayed pattern as the curve associated with superior frontal cortex showed first above-chance decoding only after 194 ms (0.504 ± 0.001), t(11) = 2.33 p = 0.0396 and rostral medial frontal as late as 365 ms.

Figure 4 with 1 supplement see all

Download asset Open asset

Only two of the chosen regions, superior frontal and superior parietal cortices, demonstrated the asymmetry between seen and unseen trials obtained with sensor-based decoding. In both cases, when trained on activity restricted to the superior frontal regions, a complete generalization obtained when trained on ‘Unseen–Correct’ but when trained on ‘Seen–Correct’ trials, decoding probability dropped for ‘Unseen–Correct’ in the fourth time window. Classification probability was significantly higher for ‘Seen–Correct’ (0.552 ± 0.017) than for the ‘Unseen–Correct’ trials (0.541 ± 0.017), t(11) = 2.35, p = 0.038. In the superior parietal cortex, when trained on ‘Unseen–Correct’ trials, asymmetry was apparent in all time windows: classifiers generalize completely for ‘Seen–Correct’ trials (Figure 4) but when trained on ‘Seen–Correct’ trials and tested on ‘Unseen–Correct’, posterior probabilities dropped. In the second window classification, probability was significantly higher for ‘Seen–Correct’ (0.534 ± 0.007) than for the ‘Unseen–Correct’ trials (0.525 ± 0.006), t(11) = 2.5, p = 0.026. In the third time window, ‘Seen–Correct’ decoding probability was higher than ‘Unseen–Correct’ ([0.537 ± 0.008] [0.523 ± 0.01], t(11) = 2.3, p = 0.036). This pattern was kept in the fourth time window, as ‘Seen–Correct’ decoding probability (0.559 ± 0.011) was higher than ‘Unseen–Correct’ decoding probability (0.545 ± 0.01), t(11) = 3.9, p = 0.002.

These findings go beyond the distinction of conscious and unconscious processing and allow us to demonstrate a representation of stimulus location outside visual areas. Above-chance classification could be achieved in all regions, including superior parietal, rostral middle frontal, and superior frontal regions, supporting the prior finding that these regions contain retinotopic maps (Sereno et al., 1995; Hagler and Sereno, 2006). Importantly, however, our results show that there is no specific region uniquely dedicated to the encoding of consciously perceived information. Instead, a better encoding on seen trials is obtained in the same overall sectors of the superior parietal cortex and superior frontal cortex that also contain location information on unseen–correct trials (i.e., blindsight trials). Again, the key difference is asymmetrical generalization, indicating the presence of superior encoding of spatial information on seen trials, only in dorsal parietal and frontal regions. This pattern primarily occurs in the late time window (beyond 270 ms), but it is already present at early stages of processing in the superior parietal cortex.

Generalization across time

According to the GNW model, conscious access corresponds to an amplification and broadcasting of the selected information. Conscious processing is, therefore, predicted to consist in a series of information-processing stages, each associated with the settling of brain activity into a temporary metastable state (stable for a duration of ∼100–300 ms), which allows the information to be flexibly transmitted to the appropriate specialized processors (Dehaene et al., 2014). The series of processing stages may differ as a function of task demands and may include verbal report (Frässle et al., 2014), executive control (Van Gaal et al., 2010), or meta-cognitive evaluation (Charles et al., 2014). On unconscious trials, the accumulation of incoming activation would fail to attain the critical threshold level needed to trigger such a discrete series of stages (‘failed ignition’), and the unconscious information would, therefore, decay over a period of a few hundreds of milliseconds (Dehaene and Naccache, 2001).

To evaluate those predictions concerning the dynamics and transient stability of internal codes, we used the temporal generalization method (King and Dehaene, 2014). This method consists in probing if a classifier trained at a certain time point t can generalize to other time points, t′. If the representation is stable, the classifier should remain efficient even if applied at a different latency. If, however, the information is successively re-encoded in a series of different brain systems, then we should see a failure of generalization beyond a certain temporal duration (i.e., away from the diagonal where t = t′) (King and Dehaene, 2014). We, therefore, quantified the endurance of conscious and unconscious representations by training and testing classifiers on all pairs of time samples (t,t′). To measure a classifier's durability independently of classification efficacy, we defined Classification Endurance (CE), as the number of samples forward and backward in time, for which decoding performance remained above 50% of its level at the original training time (t = t′)—that is, the decoder's half-life time.

Matrices of temporal generalization (Figure 5) indicated that, up to 272 ms, decoding was narrowly restricted to the diagonal for both seen and unseen trials, suggesting a fast-changing chain of perceptual processes (King and Dehaene, 2014) consistent with a feedforward propagation of unconscious location information. Beyond this point and up to ∼800 ms post stimulus, classifiers generalized to a wider neighborhood of latencies, consistent with the entry of information into evidence accumulation systems with longer time constants. The generalization pattern was examined using an analysis of variance conducted on CE in the 272–800-ms time window with factors of Visibility (Seen/Unseen), Direction (Forward/Backward in time), and Timeframe (5 successive time windows) (see Figure 5—figure supplement 1 for statistics). All main effects and their interactions were significant. On average, CE was larger for generalization forward than for generalization backward and also increased in time. We were mainly interested in the different generalization patterns of the ‘Seen–Correct’ and ‘Unseen–Correct’ trials. Surprisingly, generalization was more extended in time for unseen trials. Forward generalization showed a larger CE in unseen trials compared to seen trials (mean = 382 ms vs 201 ms, p < 0.005), and the same pattern was apparent for backward generalization to a lesser extent (mean = 159 ms vs 78 ms, p < 0.05) (see Figure 5—figure supplement 2 for statistics). Furthermore, in the forward direction, these differences interacted with time. Forward CE was constant in seen trials, while for the unseen trials, it was linearly decreasing with time. Backward CE did not differ for seen and unseen trials, but increased steadily as a function of time (see Figure 5—figure supplement 3 for statistics).

Figure 5 with 3 supplements see all

Download asset Open asset

Figure 5C shows a possible theoretical interpretation consistent with those results. The high level of decoding observed on seen trials, over a long time window of ∼270–800 ms, combined with a low level of off-diagonal temporal generalization beyond a fixed horizon of ∼160 ms, implies that a conscious location is successively encoded by a series of about four distinct stages, each terminating relatively sharply after a processing duration of ∼160 ms. Conversely, the lower decoding level and longer endurance observed for unconsciously perceived stimuli indicate that they do not enter into the same processing chain. The square pattern of the temporal generalization matrix on unseen trials (Figure 5A) suggests that unconscious information lingers in a single representational state, which decays slowly over time, thus yielding a paradoxically longer endurance than on seen trials in the forward direction. All of these findings are consistent with the above predictions as well as with earlier decoding results indicating a chain of additional processing stages unique to conscious trials (Charles et al., 2014). They suggest that conscious access can be described as the entry of information into a dynamic routing system (Sackur and Dehaene, 2009; Dehaene et al., 2014) in which information is stabilized, transferred to a processor, exploited in a determined time, and then transferred again to the next stage. As described in accumulation-of-evidence models of decision making (Shadlen and Kiani, 2011), a failure to attain a threshold-level activation at one of these stages cuts the process short, prevents the attainment of a conscious-level representation, and causes the information to linger and decay.

Previous experiments exploring the signatures of conscious access have primarily contrasted the overall brain activity evoked by perceived and unperceived stimuli. However, this comparison may be contaminated by non-specific attention, alerting, performance, and reporting confounds. The quest for brain mechanisms of consciousness implies an isolation of neural populations that encodes the details of subjective experience (Haynes, 2009). As an effort in this direction, we applied multivariate decoding techniques in order to track the neural encoding of a conscious or unconscious representation of stimulus location through time. To control over stimulus and performance confounds, we capitalized on the blindsight phenomenon and selectively analyzed a large fraction of trials with identical stimuli and accurate responses, which differed only in subjective reports of seeing or not seeing the stimulus.

The results revealed that (1) the location of a briefly flashed stimulus can be accurately decoded from MEG and EEG signals for up to 800 ms, whether or not the stimulus is seen. This finding is coherent with the existence of multiple spatial maps in occipital, parietal, and frontal cortex (Sereno et al., 1995; Hagler and Sereno, 2006). (2) Seen and unseen stimuli are initially encoded identically, but after ∼270 ms, the information is selectively amplified on ‘seen’ trials. This observation replicates and extends earlier observations with seen and unseen words, digits and pictures during masking (Del Cul et al., 2007; Fisch et al., 2009), and the attentional blink (Sergent et al., 2005). (3) Asymmetrical generalization indicates that shared spatial codes are active on unconscious and conscious trials, but that conscious trials also contain additional neural codes for stimulus location that are absent or weaker on unconscious trials. Source localization traces those conscious codes to superior parietal and superior frontal cortex. While the present method does not allow us to decide whether the difference is quantitative (the same codes are activated more strongly) or qualitative (additional neural codes are recruited), they are convergent with the prior observation that, of the many maps for space in the cortex, superior parietal and frontal maps are selectively amplified by attention, while earlier retinotopic maps are activated in an automatic manner (Saygin and Sereno, 2008). (4) The dynamics of cortical coding, as revealed by temporal generalization matrices (King and Dehaene, 2014), also differs on seen and unseen trials: conscious information undergoes a series of representational transformations, such that each decoder fails to generalize beyond ∼160 ms, while unconscious information, although weaker, generalizes over a longer period.

The latter finding fits with earlier observations that only conscious stimuli can be passed through a series of discrete information-processing stages (Sackur and Dehaene, 2009; De Lange et al., 2011; Charles et al., 2014). The observed characteristic peak-to-end duration of about 160 ms is also compatible with the finding that conscious percepts tend to be locked to an ongoing theta (∼4–7 Hz) or high delta (∼1–3 Hz) rhythm (Melloni et al., 2007; Doesburg et al., 2009; Nakatani et al., 2014; Sitt et al., 2014), which is also the frequency range of slow positive event-related potentials, such as the P3. Doesburg et al. found that perceptual switching in a binocular rivalry paradigm corresponds to the emergence of synchronized gamma rhythm embedded in theta envelope (Doesburg et al., 2009). In an attentional blink paradigm, the synchrony of fast oscillations (beta and gamma) with slow activity (theta and high delta) increased with practice, in parallel to improved stimulus visibility (Nakatani et al., 2014). Theta activity was also shown to be higher on visible trials compared to invisible trials in a masking paradigm (Melloni et al., 2007) and was successfully used to classify conscious vs vegetative patients (Sitt et al., 2014). It thus seems plausible that conscious information is able to enter into a series of computational stages that are betrayed by a series of theta- or delta-like peaks (Dehaene and Sigman, 2012). Conversely, unconscious information fails to be sequentially dispatched to the successive steps of a serial task and, therefore, remains blocked within a fixed representational state, where it slowly decays. Behavioral priming studies confirm that unconscious information lingers at detectable levels for hundreds of milliseconds (Greenwald et al., 1996) or even seconds (Soto et al., 2011).

Our results are incompatible with theories that postulate identical codes for conscious and unconscious information, which would only be distinguished by second-order metacognitive tags (Lau and Passingham, 2006). They fit with theories that postulate that a conscious episode is distinguished by an amplification of incoming sensory information, possibly through reverberating loops (Lamme, 2006), and its distributed representation in multiple distinct regions including parietal and prefrontal cortices (Dehaene and Naccache, 2001; Dehaene et al., 2006). These results are highly convergent with intracranial local-field potentials and single-neuron recordings during binocular rivalry and continuous flash suppression (Sheinberg and Logothetis, 1997; Kreiman et al., 2002; Panagiotaropoulos et al., 2012), which indicate that late neural discharges in higher cortical areas reflect which of two rivaling images are subjectively perceived. In the case of object identity, decodable neural activity is found in inferior and anterior temporal cortex (Sheinberg and Logothetis, 1997; Kreiman et al., 2002) but also in prefrontal cortex (Panagiotaropoulos et al., 2012) as reported here. Thus, this region appears at the confluence of conscious perception of object identity and location, making it a primary candidate for the integrated perception of a unified conscious scene or event (Oizumi et al., 2014).

While numerous previous studies have compared brain activity evoked by ‘seen’ and ‘unseen’ stimuli, the advantage of the present study is to specifically pursue the neural dynamics underlying a specific conscious content (stimulus location). Figuratively speaking, one may say that previous studies have extracted the ‘canvas’ on which conscious perception is painted, while our approach is aimed at extracting the ‘painting’ itself. Nevertheless, this approach suffers from several limitations. One of them is the low level of decodability achieved: single-trial EEG and MEG responses have a low signal-to-noise ratio, which limits our inferences and would not afford, for instance, a reliable brain–computer interface. Another limit of our design is that it reduces the rich experience associated with conscious perception to a single feature of a very lean stimulus, presented at threshold, which is the sole focus of attention and which is reported on every trial. Whether the present results would generalize to broader real-life situations, especially in the absence of an overt report is an important question for further research. Some recent experiments suggest that removing the need to report radically reduces the neurophysiological correlates of conscious perception, including the P3b component of event-related potentials, leaving only a posterior mid-latency component (Pitts et al., 2012, 2014a, 2014b; Frässle et al., 2014). However, many of those experiments involved a dual task, which is known to delay and dilute late ERPs (Sigman and Dehaene, 2008) and to distort conscious access (Marti et al., 2010). It might be preferable to adopt a ‘passive attentive’ paradigm involving mere attention without report. In such an auditory paradigm, late ERPs including the P3b component were again found to contain decodable information about the conscious percept (Wacongne et al., 2011; King et al., 2013). A correlation of late neurophysiological activity with perceptibility was also obtained in a variety of neurophysiological studies of human visual perception (Quiroga et al., 2008; Kouider et al., 2013) or rodent tactile perception (Manita et al., 2015). With due caution, we, therefore, hypothesize that the neurophysiological correlates of conscious perception uncovered here might be generalized to other experimental conditions.

1. 1. Aru J
   2. Bachmann T
   3. Singer W
   4. Melloni L

   (2012) Distilling the neural correlates of consciousness

   Neuroscience and Biobehavioral Reviews 36:737–746.

   https://doi.org/10.1016/j.neubiorev.2011.12.003

   • Google Scholar
2. 1. Brainard DH

   (1997) The psychophysics toolbox

   Spatial Vision 10:433–436.

   https://doi.org/10.1163/156856897X00357

   • Google Scholar
3. 1. Chang CC
   2. Lin CJ

   (2011) LIBSVM: a library for support vector machines

   ACM Transactions on Intelligent Systems and Technology 2:27.

   https://doi.org/10.1145/1961189.1961199

   • Google Scholar
4. 1. Charles L
   2. King J-R
   3. Dehaene S

   (2014) Decoding the dynamics of action, intention, and error detection for conscious and subliminal stimuli

   The Journal of Neuroscience 34:1158–1170.

   https://doi.org/10.1523/JNEUROSCI.2465-13.2014

   • Google Scholar
5. 1. Cowey A
   2. Stoerig P

   (1995) Blindsight in monkeys

   Nature 373:247–249.

   https://doi.org/10.1038/373247a0

   • Google Scholar
6. 1. De Lange FP
   2. van Gaal S
   3. Lamme VA
   4. Dehaene S

   (2011) How awareness changes the relative weights of evidence during human decision-making

   PLOS Biology 9:e1001203.

   https://doi.org/10.1371/journal.pbio.1001203

   • Google Scholar
7. 1. Dehaene S
   2. Changeux JP

   (2011) Experimental and theoretical approaches to conscious processing

   Neuron 70:200–227.

   https://doi.org/10.1016/j.neuron.2011.03.018

   • Google Scholar
8. 1. Dehaene S
   2. Changeux JP
   3. Naccache L
   4. Sackur J
   5. Sergent C

   (2006) Conscious, preconscious, and subliminal processing: a testable taxonomy

   Trends in Cognitive Sciences 10:204–211.

   https://doi.org/10.1016/j.tics.2006.03.007

   • Google Scholar
9. 1. Dehaene S
   2. Charles L
   3. King J-R
   4. Marti S

   (2014) Toward a computational theory of conscious processing

   Current Opinion in Neurobiology 25:76–84.

   https://doi.org/10.1016/j.conb.2013.12.005

   • Google Scholar
10. 1. Dehaene S
    2. Naccache L

    (2001) Towards a cognitive neuroscience of consciousness: basic evidence and a workspace framework

    Cognition 79:1–37.

    https://doi.org/10.1016/S0010-0277(00)00123-2

    • Google Scholar
11. 1. Dehaene S
    2. Sigman M

    (2012) From a single decision to a multi-step algorithm

    Current Opinion in Neurobiology 22:937–945.

    https://doi.org/10.1016/j.conb.2012.05.006

    • Google Scholar
12. 1. Dehaene S
    2. Sergent C
    3. Changeux JP

    (2003) A neuronal network model linking subjective reports and objective physiological data during conscious perception

    Proceedings of the National Academy of Sciences of USA 100:8520–8525.

    https://doi.org/10.1073/pnas.1332574100

    • Google Scholar
13. 1. Del Cul A
    2. Baillet S
    3. Dehaene S

    (2007) Brain dynamics underlying the nonlinear threshold for access to consciousness

    PLOS Biology 5:e260.

    https://doi.org/10.1371/journal.pbio.0050260

    • Google Scholar
14. 1. Doesburg SM
    2. Green JJ
    3. McDonald JJ
    4. Ward LM

    (2009) Rhythms of consciousness: binocular rivalry reveals large-scale oscillatory network dynamics mediating visual perception

    PLOS One 4:e6142.

    https://doi.org/10.1371/journal.pone.0006142

    • Google Scholar
15. 1. Fahrenfort JJ
    2. Snijders TM
    3. Heinen K
    4. van Gaal S
    5. Scholte HS
    6. Lamme VA

    (2012) Neuronal integration in visual cortex elevates face category tuning to conscious face perception

    Proceedings of the National Academy of Sciences of USA 109:21504–21509.

    https://doi.org/10.1073/pnas.1207414110

    • Google Scholar
16. 1. Fisch L
    2. Privman E
    3. Ramot M
    4. Harel M
    5. Nir Y
    6. Kipervasser S
    7. Andelman F
    8. Neufeld MY
    9. Kramer U
    10. Fried I
    11. Malach R

    (2009) Neural “ignition”: enhanced activation linked to perceptual awareness in human ventral stream visual cortex

    Neuron 64:562–574.

    https://doi.org/10.1016/j.neuron.2009.11.001

    • Google Scholar
17. 1. Frässle S
    2. Sommer J
    3. Jansen A
    4. Naber M
    5. Einhäuser W

    (2014) Binocular rivalry: frontal activity relates to introspection and action but not to perception

    The Journal of Neuroscience 34:1738–1747.

    https://doi.org/10.1523/JNEUROSCI.4403-13.2014

    • Google Scholar
18. 1. Friston K

    (2005) A theory of cortical responses

    Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences 360:815–836.

    https://doi.org/10.1098/rstb.2005.1622

    • Google Scholar
19. 1. Gramfort A
    2. Luessi M
    3. Larson E
    4. Engemann D
    5. Strohmeier D
    6. Brodbeck C
    7. Parkkonen L
    8. Hämäläinen M

    (2014) MNE software for processing MEG and EEG data

    Neuroimage 86:446–460.

    https://doi.org/10.1016/j.neuroimage.2013.10.027

    • Google Scholar
20. 1. Greenwald AG
    2. Draine SC
    3. Abrams RL

    (1996) Three cognitive markers of unconscious semantic activation

    Science 273:1699–1702.

    https://doi.org/10.1126/science.273.5282.1699

    • Google Scholar
21. 1. Hagler DJ
    2. Sereno MI

    (2006) Spatial maps in frontal and prefrontal cortex

    Neuroimage 29:567–577.

    https://doi.org/10.1016/j.neuroimage.2005.08.058

    • Google Scholar
22. 1. Haynes JD
    2. Rees G

    (2006) Decoding mental states from brain activity in humans

    Nature Reviews. Neuroscience 7:523–534.

    https://doi.org/10.1038/nrn1931

    • Google Scholar
23. 1. Haynes JD

    (2009) Decoding visual consciousness from human brain signals

    Trends in Cognitive Sciences 13:194–202.

    https://doi.org/10.1016/j.tics.2009.02.004

    • Google Scholar
24. 1. Kamitani Y
    2. Tong F

    (2005) Decoding the visual and subjective contents of the human brain

    Nature Neuroscience 8:679–685.

    https://doi.org/10.1038/nn1444

    • Google Scholar
25. 1. King J-R
    2. Dehaene S

    (2014) Characterizing the dynamics of mental representations: the temporal generalization method

    Trends in Cognitive Sciences 18:203–210.

    https://doi.org/10.1016/j.tics.2014.01.002

    • Google Scholar
26. 1. King JR
    2. Gramfort A
    3. Schurger A
    4. Naccache L
    5. Dehaene S

    (2014) Two distinct dynamic modes subtend the detection of unexpected sounds

    PLOS One 9:e85791.

    https://doi.org/10.1371/journal.pone.0085791

    • Google Scholar
27. 1. King JR
    2. Faugeras F
    3. Gramfort A
    4. Schurger A
    5. El Karoui I
    6. Sitt JD
    7. Rohaut B
    8. Wacongne C
    9. Labyt E
    10. Bekinschtein T
    11. Cohen L
    12. Naccache L
    13. Dehaene S

    (2013) Single-trial decoding of auditory novelty responses facilitates the detection of residual consciousness

    Neuroimage 83:726–738.

    https://doi.org/10.1016/j.neuroimage.2013.07.013

    • Google Scholar
28. Book

    1. Koch C

    (2004)

    The quest for consciousness: a neurobiological approach

    New York: Roberts & co.

    • Google Scholar
29. 1. Kouider S
    2. Stahlhut C
    3. Gelskov SV
    4. Barbosa LS
    5. Dutat M
    6. De Gardelle V
    7. Christophe A
    8. Dehaene S
    9. Dehaene-Lambertz G

    (2013) A neural marker of perceptual consciousness in infants

    Science 340:376–380.

    https://doi.org/10.1126/science.1232509

    • Google Scholar
30. 1. Kreiman G
    2. Fried I
    3. Koch C

    (2002) Single-neuron correlates of subjective vision in the human medial temporal lobe

    Proceedings of the National Academy of Sciences of USA 99:8378–8383.

    https://doi.org/10.1073/pnas.072194099

    • Google Scholar
31. 1. Lamme VA

    (2006) Towards a true neural stance on consciousness

    Trends in Cognitive Sciences 10:494–501.

    https://doi.org/10.1016/j.tics.2006.09.001

    • Google Scholar
32. 1. Lamy D
    2. Salti M
    3. Bar-Haim Y

    (2009) Neural correlates of subjective awareness and unconscious processing: an ERP study

    Journal of Cognitive Neuroscience 21:1435–1446.

    https://doi.org/10.1162/jocn.2009.21064

    • Google Scholar
33. 1. Lau HC
    2. Passingham RE

    (2006) Relative blindsight in normal observers and the neural correlate of visual consciousness

    Proceedings of the National Academy of Sciences of USA 103:18763–18768.

    https://doi.org/10.1073/pnas.0607716103

    • Google Scholar
34. 1. Lau H
    2. Rosenthal D

    (2011) Empirical support for higher-order theories of conscious awareness

    Trends in Cognitive Sciences 15:365–373.

    https://doi.org/10.1016/j.tics.2011.05.009

    • Google Scholar
35. 1. Laureys S
    2. Owen AM
    3. Schiff ND

    (2004) Brain function in coma, vegetative state, and related disorders

    Lancet Neurology 3:537–546.

    https://doi.org/10.1016/S1474-4422(04)00852-X

    • Google Scholar
36. 1. Levitt H

    (1971) Transformed up-down methods in psychoacoustics

    The Journal of the Acoustical Society of America 49:467–477.

    https://doi.org/10.1121/1.1912375

    • Google Scholar
37. 1. Manita S
    2. Suzuki T
    3. Homma C
    4. Matsumoto T
    5. Odagawa M
    6. Yamada K
    7. Ota K
    8. Matsubara C
    9. Inutsuka A
    10. Sato M
    11. Ohkura M
    12. Yamanaka A
    13. Yanagawa Y
    14. Nakai J
    15. Hayashi Y
    16. Larkum ME
    17. Murayama M

    (2015) A top-down cortical circuit for accurate sensory perception

    Neuron 86:1304–1316.

    https://doi.org/10.1016/j.neuron.2015.05.006

    • Google Scholar
38. 1. Marti S
    2. Sackur J
    3. Sigman M
    4. Dehaene S

    (2010) Mapping introspection's blind spot: reconstruction of dual-task phenomenology using quantified introspection

    Cognition 115:303–313.

    https://doi.org/10.1016/j.cognition.2010.01.003

    • Google Scholar
39. 1. Melloni L
    2. Molina C
    3. Pena M
    4. Torres D
    5. Singer W
    6. Rodriguez E

    (2007) Synchronization of neural activity across cortical areas correlates with conscious perception

    The Journal of Neuroscience 27:2858–2865.

    https://doi.org/10.1523/JNEUROSCI.4623-06.2007

    • Google Scholar
40. 1. Nakatani C
    2. Raffone A
    3. van Leeuwen C

    (2014) Efficiency of conscious access improves with coupling of slow and fast neural oscillations

    Journal of Cognitive Neuroscience 26:1168–1179.

    https://doi.org/10.1162/jocn_a_00540

    • Google Scholar
41. 1. Nieuwenhuis S
    2. Gilzenrat MS
    3. Holmes BD
    4. Cohen JD

    (2005) The role of the locus coeruleus in mediating the attentional blink: a neurocomputational theory

    Journal of Experimental Psychology. General 134:291–307.

    https://doi.org/10.1037/0096-3445.134.3.291

    • Google Scholar
42. 1. Oizumi M
    2. Albantakis L
    3. Tononi G

    (2014) From the phenomenology to the mechanisms of consciousness: integrated information theory 3.0

    PLOS Computational Biology 10:e1003588.

    https://doi.org/10.1371/journal.pcbi.1003588

    • Google Scholar
43. 1. Oostenveld R
    2. Fries P
    3. Maris E
    4. Schoffelen JM

    (2011) FieldTrip:open source software for advanced analysis of MEG, EEG, and invasive electrophysiological data

    Computational Intelligence and Neuroscience 2011:156869.

    https://doi.org/10.1155/2011/156869

    • Google Scholar
44. 1. Panagiotaropoulos TI
    2. Deco G
    3. Kapoor V
    4. Logothetis NK

    (2012) Neuronal discharges and gamma oscillations explicitly reflect visual consciousness in the lateral prefrontal cortex

    Neuron 74:924–935.

    https://doi.org/10.1016/j.neuron.2012.04.013

    • Google Scholar
45. 1. Pitts MA
    2. Martinez A
    3. Hillyard SA

    (2012) Visual processing of contour patterns under conditions of inattentional blindness

    Journal of Cognitive Neuroscience 24:287–303.

    https://doi.org/10.1162/jocn_a_00111

    • Google Scholar
46. 1. Pitts MA
    2. Metzler S
    3. Hillyard SA

    (2014a) Isolating neural correlates of conscious perception from neural correlates of reporting one's perception

    Frontiers in Psychology 5:1078.

    https://doi.org/10.3389/fpsyg.2014.01078

    • Google Scholar
47. 1. Pitts MA
    2. Padwal J
    3. Fennelly D
    4. Martínez A
    5. Hillyard SA

    (2014b) Gamma band activity and the P3 reflect post-perceptual processes, not visual awareness

    Neuroimage 101:337–350.

    https://doi.org/10.1016/j.neuroimage.2014.07.024

    • Google Scholar
48. 1. Platt JC

    (1999)

    Advances in kernel methods

    185–208, Advances in kernel methods, MIT press.

    • Google Scholar
49. 1. Quiroga RQ
    2. Mukamel R
    3. Isham EA
    4. Malach R
    5. Fried I

    (2008) Human single-neuron responses at the threshold of conscious recognition

    Proceedings of the National Academy of Sciences of USA 105:3599–3604.

    https://doi.org/10.1073/pnas.0707043105

    • Google Scholar
50. 1. Rossetti Y

    (1998) Implicit short-lived motor representations of space in brain damaged and healthy subjects

    Consciousness and Cognition 7:520–558.

    https://doi.org/10.1006/ccog.1998.0370

    • Google Scholar
51. 1. Sackur J
    2. Dehaene S

    (2009) The cognitive architecture for chaining of two mental operations

    Cognition 111:187–211.

    https://doi.org/10.1016/j.cognition.2009.01.010

    • Google Scholar
52. 1. Saygin AP
    2. Sereno MI

    (2008) Retinotopy and attention in human occipital, temporal, parietal, and frontal cortex

    Cerebral Cortex 18:2158–2168.

    https://doi.org/10.1093/cercor/bhm242

    • Google Scholar
53. 1. Sereno MI
    2. Dale AM
    3. Reppas JB
    4. Kwong KK
    5. Belliveau JW
    6. Brady TJ
    7. Rosen BR
    8. Tootell RB

    (1995) Borders of multiple visual areas in humans revealed by functional magnetic resonance imaging

    Science 268:889–893.

    https://doi.org/10.1126/science.7754376

    • Google Scholar
54. 1. Sergent C
    2. Baillet S
    3. Dehaene S

    (2005) Timing of the brain events underlying access to consciousness during the attentional blink

    Nature Neuroscience 8:1391–1400.

    https://doi.org/10.1038/nn1549

    • Google Scholar
55. Book

    1. Shadlen MN
    2. Kiani R

    (2011)

    Consciousness as a decision to engage

    In: Dehaene S, Christen Y, editors. Characterizing Consciousness: From Cognition to the Clinic?. Berlin Heidelberg: Springer Verlag. pp. 27–46.

    • Google Scholar
56. 1. Sheinberg DL
    2. Logothetis NK

    (1997) The role of temporal cortical areas in perceptual organization

    Proceedings of the National Academy of Sciences of USA 94:3408–3413.

    https://doi.org/10.1073/pnas.94.7.3408

    • Google Scholar
57. 1. Sigman M
    2. Dehaene S

    (2008) Brain mechanisms of serial and parallel processing during dual-task performance

    The Journal of Neuroscience 28:7585–7598.

    https://doi.org/10.1523/JNEUROSCI.0948-08.2008

    • Google Scholar
58. 1. Sitt JD
    2. King J-R
    3. El Karoui I
    4. Rohaut B
    5. Faugeras F
    6. Gramfort A
    7. Cohen L
    8. Sigman M
    9. Dehaene S
    10. Naccache L

    (2014) Large scale screening of neural signatures of consciousness in patients in a vegetative or minimally conscious state

    Brain 137:2258–2270.

    https://doi.org/10.1093/brain/awu141

    • Google Scholar
59. 1. Soto D
    2. Mantyla T
    3. Silvanto J

    (2011) Working memory without consciousness

    Current Biology 21:R912–R913.

    https://doi.org/10.1016/j.cub.2011.09.049

    • Google Scholar
60. 1. Taulu S
    2. Kajola M
    3. Simola J

    (2004) Suppression of interference and artifacts by the signal space separation method

    Brain Topography 16:269–275.

    https://doi.org/10.1023/B:BRAT.0000032864.93890.f9

    • Google Scholar
61. 1. Van Gaal S
    2. Lamme VA
    3. Fahrenfort JJ
    4. Ridderinkhof KR

    (2010) Dissociable brain mechanisms underlying the conscious and unconscious control of behavior

    Journal of Cognitive Neuroscience 23:91–105.

    https://doi.org/10.1162/jocn.2010.21431

    • Google Scholar
62. 1. Wacongne C
    2. Labyt E
    3. van Wassenhove V
    4. Bekinschtein T
    5. Naccache L
    6. Dehaene S

    (2011) Evidence for a hierarchy of predictions and prediction errors in human cortex

    Proceedings of the National Academy of Sciences of USA 108:20754–20759.

    https://doi.org/10.1073/pnas.1117807108

    • Google Scholar
63. 1. Weiskrantz L

    (1996) Blindsight revisited

    Current Opinion in Neurobiology 6:215–220.

    https://doi.org/10.1016/S0959-4388(96)80075-4

    • Google Scholar

Thank you for sending your work entitled “Distinct cortical codes and temporal dynamics for conscious and unconscious percepts” for consideration at eLife. Your article has been favorably evaluated by Eve Marder (Senior editor), two reviewers, and Heidi Johansen-Berg, who is a member of our Board of Reviewing Editors.

The Reviewing editor and the reviewers discussed their comments before we reached this decision, and the Reviewing editor has assembled the following comments to help you prepare a revised submission.

All reviewers found the study well designed and carefully analysed. All reviewers felt the paper was a valuable addition to the literature and offered several advantages relative to prior work in this field. The following specific issues raised by reviewers should be addressed in a revision:

Essential points:

1) It would be useful to relate the decoding strategies used here to previous studies that used more traditional approaches to data analysis. For example, the authors could include a figure in which ERFs and ERPs are compared across the three main conditions (seen-correct, unseen-correct, unseen-incorrect). Specifically, it is worth reporting whether the difference between seen and unseen decoding accuracies observed in the late time period (<270 ms) corresponds to the N2-P3 waves reported in previous studies.

2) Unless we are missing something, the authors did not run any analyses in which multivariate decoding was used to classify seen vs. unseen responses. Of the many analyses reported, a difference in accuracy in decoding stimulus location was found for seen vs. unseen trials, but both were above chance. Was a classifier ever directly used to discriminate seen vs. unseen trials?

3) Perhaps the late activity involves online conscious control of the speeded hand-movement response rather than conscious perception? We recognize that the authors have included a control in which no stimulus is present and responses are made. However, that control does not fully address the potential concern. Would the authors predict similar late-stage decoding of stimulus location if a 500 ms stimulus was presented (so as to be obviously seen) and subjects were not required to make any responses? This requires additional discussion.

4) It is not clear from the Methods section how EEG and MEG data were separately analyzed or combined in the multivariate decoding. For example, Figure 2A shows “decoding location from MEG/EEG”, implying that signals from both were utilized in the decoding. More details regarding how these signals were separately analyzed or combined should be provided in the Methods section.

5) The authors should report reaction times for the speeded location identification response.

The authors may choose to comment on the following discussion points in a revision:

1) The authors make several well-informed predictions based on current theories or previous evidence, and the data appear to support these predictions. However, strong claims are often made based on very small effects. The authors should acknowledge this and should comment on the need for caution in interpreting a 1% difference in decoding accuracy between seen versus unseen trials. It would also be helpful to comment on why decoding accuracies were only just barely (though significantly) above chance to begin with.

2) If the late activity (270-800 ms) corresponds to conscious perception, the authors should consider explaining why the neural events underlying the conscious experience of a 33 ms line stimulus last 16 times longer (530 ms). Isn't it more intuitive that the conscious experience (and associated neural events) of a very brief stimulus would be correspondingly very brief?

3) Please provide more discussion of the finding that conscious representations seem to have a fixed duration of 160 ms which then gives rise to a new cycle, while unconscious representations decay slowly. If the authors could discuss this phenomena more extensively (e.g., do neural oscillations play a role? Does it reflect engagement of successive brain areas?), it would shed light on the mechanism behind this novel and interesting finding.

1) It would be useful to relate the decoding strategies used here to previous studies that used more traditional approaches to data analysis. For example, the authors could include a figure in which ERFs and ERPs are compared across the three main conditions (seen-correct, unseen-correct, unseen-incorrect). Specifically, it is worth reporting whether the difference between seen and unseen decoding accuracies observed in the late time period (<270ms) corresponds to the N2-P3 waves reported in previous studies.

The objective of our study was to go beyond existing data on the neural correlates of conscious perception, specifically by attempting to decode the neural dynamics of a specific content, both when it is consciously perceived and when it is processed non-consciously. This is the reason why we did not initially include the basic ERFs and ERPs for the three conditions. We agree, however, that this is important background information. Therefore, we have now included this information in the paper. The results indeed demonstrate classic differences in the N2 and P3 waves in ERPs, and the corresponding events in ERFs. We have added a specific subsection on this topic to the Results section (“Relation to previous findings from event-related responses”).

“Our fourth time window, where significantly improved location decoding was observed on seen compared to unseen trials […] the latter corresponds to a search for brain states encoding a specific consciously perceived content.”

2) Unless we are missing something, the authors did not run any analyses in which multivariate decoding was used to classify seen vs. unseen responses. Of the many analyses reported, a difference in accuracy in decoding stimulus location was found for seen vs. unseen trials, but both were above chance. Was a classifier ever directly used to discriminate seen vs. unseen trials?

Again, this was not the primary goal of our study. Numerous previous studies have studied the neural signatures of “Seen” and “Unseen” stimuli, but they were not able to identify the neural activity directly related to consciously perceived contents. Here we tried to pursue the neural dynamics underlying consciously perceived contents. Figuratively speaking, the old approaches have extracted the “canvas” on which conscious perception was painted while our approach is aimed at extracting the “painting” itself. We agree, however, that it is also useful to have additional information on the capacity to decode seen and unseen trials from M/EEG data. This information is now included in the paper (in the subsection headed “Relation to previous findings from event-related responses”):

”Training a classifier to use those event-related responses in order to separate “Seen-Correct” versus “Unseen-Correct” trials yielded a modest but significant classification success […]. This reduced the number of trials on which decoder was trained on and affected classification results.”

3) Perhaps the late activity involves online conscious control of the speeded hand-movement response rather than conscious perception? We recognize that the authors have included a control in which no stimulus is present and responses are made. However, that control does not fully address the potential concern. Would the authors predict similar late-stage decoding of stimulus location if a 500ms stimulus was presented (so as to be obviously seen) and subjects were not required to make any responses? This requires additional discussion.

Two valid points were raised here. The first point is the hypothesis that the late stage of the classification could reflect conscious response planning. We believe that our control, decoding target-absent trials, does address this concern, but we have added two arguments to support our claim that the decoding difference is related to conscious perception and not to conscious motor preparation (see subsection headed “Controls for hand and eye movements”). The revised paragraph now reads:

“Although it remains possible that conscious response planning modulates the late stages of location processing, several aspects of our experiments argue against a contribution of motor codes to our results. […] Nor can the improved classification performance for “Seen-Correct” over “Unseen-Correct”, since those trials involve exactly the same stimuli and responses.”

The second point concerns the generality of our results outside the specific experimental environment and task. In real-life situations, conscious contents are richer, continuous and do not necessarily require a report. The experimental design here used a very lean stimulus. Only one of its features (location) was of relevance to the participants, who had to explicitly report it and its visibility. Generalization beyond the experimental situation should consider whether the dynamics revealed here could be generalized to other features of the same stimuli, to other stimuli and to other tasks, including passive viewing. Here we set a methodological and theoretical framework and establish preliminary results. Obviously, a lot of additional work is needed to address these points of concern. We have added a final paragraph to the Discussion where we discuss the limits of our findings and argue, cautiously, that similar dynamics would be found if we had used clearly visible stimuli without any overt response. We base this conclusion on parallels with other experiments, for instance using the auditory local-global paradigm, where similar findings of a late activity correlated with conscious perception was observed even in the absence of any overt report. We do note, however, that there is a current debate in the literature concerning this very issue. A discussion of those points has been added in the final paragraph:

“While numerous previous studies have compared brain activity evoked by “seen” and “unseen” stimuli, the advantage of the present study is to specifically pursue the neural dynamics underlying a specific conscious content (stimulus location). […] With due caution, we therefore hypothesize that the neurophysiological correlates of conscious perception uncovered here might generalized to other experimental conditions.”

4) It is not clear from the Methods section how EEG and MEG data were separately analyzed or combined in the multivariate decoding. For example, Figure 2A shows “decoding location from MEG/EEG”, implying that signals from both were utilized in the decoding. More details regarding how these signals were separately analyzed or combined should be provided in the Methods section.

All of our decoding was based on combined MEG and EEG signals. We have added a precise description of our methods in the subsection “Preprocessing” of the Materials and methods:

“…To combine data from MEG magnetometer and planar gradiometer as well as from EEG channels, a normalization procedure was first applied; the baseline standard deviation was estimated for each channel using all the trials in the experiment and then all the samples were divided by this std to yield a z-score which was entered in the classification algorithm”.

5) The authors should report reaction times for the speeded location identification response.

We reported the reaction times in the previous version in the Materials and methods section. We now moved it to the Results section:

“Mean localization RTs were 812.2±44 ms. Localization RTs for different critical conditions were as follow: 701.4±35 ms for “Seen-Correct” trials, 878.1±78 ms for “Unseen-Correct” trials and 1110±101 ms for “Unseen-incorrect” trials.”

The authors may choose to comment on the following discussion points in a revision:

1) The authors make several well-informed predictions based on current theories or previous evidence, and the data appear to support these predictions. However, strong claims are often made based on very small effects. The authors should acknowledge this and should comment on the need for caution in interpreting a 1% difference in decoding accuracy between seen versus unseen trials. It would also be helpful to comment on why decoding accuracies were only just barely (though significantly) above chance to begin with.

We agree and have toned down our claims whenever possible, particularly in an additional final paragraph of the Discussion, pointing to the limitations of our study. It is important to note that we are decoding single-trial responses. It is well-known that the signal-to-noise ratio in EEG and MEG is quite low, owing to a superimposition of ongoing spontaneous activity whose intensity can be more than ten times larger than evoked activity. Thus, the absolute decoding level is necessarily low. Our argument is not based on an absolute level of decoding (unlike, say, experiments that aim at efficient brain–computer interfaces), but on the mere presence of differentially decodable information, even if it remains at a low level.

2) If the late activity (270-800ms) corresponds to conscious perception, the authors should consider explaining why the neural events underlying the conscious experience of a 33ms line stimulus last 16 times longer (530ms). Isn't it more intuitive that the conscious experience (and associated neural events) of a very brief stimulus would be correspondingly very brief?

We disagree with this intuition. Although the stimulus is brief, its internal representation may be long-lasting. There is considerable evidence that even a brief sensory event may cause an entire series of brain activations, lasting way beyond the duration of the stimulus. The vast majority of theories of consciousness, including Global Neuronal Workspace theory, acknowledge that there is a tight relation between conscious perception and working memory. Even briefly presented stimuli, if conscious, must remain efficiently coded and “meta-stable” for a duration long enough to allow reportability long after the stimulus is gone. Indeed, a direct relation between conscious perception and long-lasting metastability has been demonstrated in previous research. See for instance the following two papers:

King, J.-R., Gramfort, A., Schurger, A., Naccache, L., & Dehaene, S. (2014). Two distinct dynamic modes subtend the detection of unexpected sounds. PloS One, 9(1), e85791. http://doi.org/10.1371/journal.pone.0085791

Schurger, A., Sarigiannidis, I., Naccache, L., Sitt, J. D., & Dehaene, S. (2015). Cortical activity is more stable when sensory stimuli are consciously perceived. Proceedings of the National Academy of Sciences of the United States of America, 112(16), E2083–2092. http://doi.org/10.1073/pnas.1418730112

3) Please provide more discussion of the finding that conscious representations seem to have a fixed duration of 160ms which then gives rise to a new cycle, while unconscious representations decay slowly. If the authors could discuss this phenomena more extensively (e.g., do neural oscillations play a role? Does it reflect engagement of successive brain areas?), it would shed light on the mechanism behind this novel and interesting finding.

We agree that this rough duration of 160ms is a very interesting finding. We have added some discussion of its potential relation to theta frequencies, as the observed duration of successive conscious stages is compatible with a nearly rhythmic succession of processing stages at this frequency. We now cite several publications that have demonstrated a tight correlation between theta oscillations and conscious perception. A discussion of this point has been added to the Discussion:

“The observed characteristic peak-to-end duration of about 160ms is also compatible with the finding that conscious percepts tend to be locked to an ongoing theta (∼4–7 Hz) or high delta (∼1–3 Hz) rhythm […]. It thus seems plausible that conscious information is able to enter into a series of computational stages that are betrayed by a series of theta- or delta-like peaks (Dehaene and Sigman, 2012).”

Author details

1. Moti Salti

   1. Cognitive Neuroimaging Unit, Institut National de la Santé et de la Recherche Médicale, Gif sur Yvette, France
   2. Neurospin center, Institut d'Imagerie Biomédicale, Gif sur Yvette, France
   3. Department of Brain and Cognitive Sciences, Ben-Gurion University of the Negev, Beer-Sheva, Israel
   4. Zlotowski Center for Neuroscience, Ben-Gurion University of the Negev, Beer-Sheva, Israel

   Contribution

   MS, Conception and design, Acquisition of data, Analysis and interpretation of data, Drafting or revising the article

   For correspondence

   motisalti@gmail.com

   Competing interests

   The authors declare that no competing interests exist.

2. Simo Monto

   1. Cognitive Neuroimaging Unit, Institut National de la Santé et de la Recherche Médicale, Gif sur Yvette, France
   2. Neurospin center, Institut d'Imagerie Biomédicale, Gif sur Yvette, France
   3. Department of Neuroscience and Biomedical Engineering, Aalto University School of Science, Espoo, Finland

   Contribution

   SM, Analysis and interpretation of data, Drafting or revising the article

   Competing interests

   The authors declare that no competing interests exist.

3. Lucie Charles

   1. Cognitive Neuroimaging Unit, Institut National de la Santé et de la Recherche Médicale, Gif sur Yvette, France
   2. Neurospin center, Institut d'Imagerie Biomédicale, Gif sur Yvette, France
   3. Attention and Cognitive Control Lab Department of Experimental Psychology, University of Oxford, Oxford, United Kingdom

   Contribution

   LC, Acquisition of data, Drafting or revising the article

   Competing interests

   The authors declare that no competing interests exist.

4. Jean-Remi King

   1. Cognitive Neuroimaging Unit, Institut National de la Santé et de la Recherche Médicale, Gif sur Yvette, France
   2. Neurospin center, Institut d'Imagerie Biomédicale, Gif sur Yvette, France

   Contribution

   J-RK, Analysis and interpretation of data, Drafting or revising the article

   Competing interests

   The authors declare that no competing interests exist.

5. Lauri Parkkonen

   1. Cognitive Neuroimaging Unit, Institut National de la Santé et de la Recherche Médicale, Gif sur Yvette, France
   2. Neurospin center, Institut d'Imagerie Biomédicale, Gif sur Yvette, France
   3. Department of Neuroscience and Biomedical Engineering, Aalto University School of Science, Espoo, Finland

   Contribution

   LP, Conception and design, Analysis and interpretation of data, Drafting or revising the article

   Competing interests

   The authors declare that no competing interests exist.

6. Stanislas Dehaene

   1. Cognitive Neuroimaging Unit, Institut National de la Santé et de la Recherche Médicale, Gif sur Yvette, France
   2. Neurospin center, Institut d'Imagerie Biomédicale, Gif sur Yvette, France
   3. Collège de France, Paris, France
   4. University of Paris-Sud, Orsay, France

   Contribution

   SD, Conception and design, Analysis and interpretation of data, Drafting or revising the article

   Competing interests

   The authors declare that no competing interests exist.

• 3,492

  Page views

• 973

  Downloads

• 61

  Citations

Article citation count generated by polling the highest count across the following sources: Scopus, Crossref, PubMed Central.