Some of the propositions summarized in the following chapter have been derived from experiments on the neuronal substrate of consciousness that have been described in detail in Melloni & Singer (2011) and Aru et al. (2012a, 2012b). A few decades ago, attempts to identify the neuronal correlates of consciousness (NCC) were considered futile. Because of the rapid development of non-invasive technology for the registration of neuronal activity in the human brain, and because of advances in the analysis of the neuronal underpinnings of higher cognitive functions, the search for NCC has now become a very active field of research in cognitive neuroscience. As expected, this new field is confronted with great challenges that are difficult to overcome. The explanandum is ill-defined; the prerequisites and consequences of conscious processing cannot easily be distinguished by experiment from conscious processing per se; the experience of mental causation and agency is difficult to reconcile with contemporary concepts of self-organization; and finally it is difficult to bridge the epistemic gap between phenomena that are experienced from a first-person perspective and mechanisms described from a third-person perspective. Some of these problems will be addressed in the following paragraphs.
Most languages have coined a term for consciousness. Thus, it must be a robust phenomenon on which human beings can agree. However, while it is easy to use the term, it is virtually impossible to give a formal definition of what exactly it means. Nevertheless, the implicit understanding of what it is to be conscious seems to be sufficiently clear and widely accepted enough to justify a search for its neuronal correlates and, ultimately, to identify the neuronal mechanisms that enable a subject to be conscious of something. In their seminal paper, Crick & Koch (1990) propose that consciousness is a specific cognitive function and that, as such, it must have neuronal correlates that can be analyzed with the tools of the natural sciences. With the development of non-invasive imaging technologies, the tools became available to actually pursue this project and the search for the neuronal correlates of consciousness (NCC) became a mainstream endeavour.
Before discussing some of the proposed theories for NCC, I shall attempt to give an operational definition of what I mean when referring to awareness and consciousness or, in other terms, what it means to be aware of something or to be conscious. Subjects will be considered aware of something if they are able to report the presence or absence of the content of a cognitive process—irrespective of whether this content is made available by recall from stored memories or drawn from actual sensory experience. Thus, one criterion for awareness is the reportability of the presence of a cognitive content. These reports can in principle consist of any motor response, but to be on the safe side, it is often requested that the report be verbal. The reason for this is that behavioural responses can be obtained under forced choice conditions that clearly indicate that the brain has processed and recognized the respective sensory material and produced a correct response even though the subject may not have been aware of having perceived the stimulus. There is thus an inherent ambiguity in non-verbal responses. They can but need not necessarily signal awareness, and this constrains research on NCC in animal experiments. Since consciousness is so difficult to define, an attempt will be made to avoid this term. Instead we shall use the adverb “consciously” and the adjective “conscious” in order to further specify particular brain states or aspects of a perceptual process. In addition, research into the hard problem of consciousness (Chalmers 2000) confronts the problem of explaining the phase transition from neuronal processes to the qualia of subjective experience, but this will be discussed only briefly at the end of this paper—and there in an enlarged context that transcends neurobiological approaches by also taking social interactions into account.
The state of being aware of something has a number of distinct properties that constrain the underlying neuronal mechanisms. One important feature of this state is unity or relatedness: Contents of which one is aware are experienced as simultaneously present and related to each other. Because of the distributed organization of brain processes, mechanisms supporting phenomenal awareness must therefore be able to bind together computational results obtained in multiple specialized and widely distributed processing areas. Another feature of awareness is that the contents that one is aware of change continuously but are bound together in time, appearing as a seamless flow that is coherent in space and time. Finally, subjects are only aware of a small fraction of on-going cognitive operations. Still, even signals of which subjects are not aware are often readily processed and impact behaviour (Dehaene et al. 1998; van Gaal et al. 2008). Thus there must be gating mechanism that determines which signals are processed consciously, which are processed and control behaviour but remain unconscious, and which are excluded from processing altogether. Therefore, the identification of NCC requires a clear delineation between subconscious and conscious processes and an analysis of the mechanisms that gate access to awareness.
As mentioned above, an enormous amount of knowledge is stored in the specific architectures of the brain, but we are not aware of most of these “given” heuristics, assumptions, and concepts. These routines determine the outcome of cognitive processes, which often have access to conscious recollection while themselves remaining hidden in the unconscious. We cannot move these implicit hypotheses and rules to the workspace of consciousness by focusing our attention on them, as is possible with most sensory signals and contents stored, for example, in “declarative memory”—the memory in which is inscribed what has been consciously experienced. Excluded from conscious experience are also certain sensory signals—such as those elicited, for example, by pheromones, which are processed by special olfactory subsystems—or the many signals from within the body—such as messages about blood pressures, sugar levels, and so on. It cannot be emphasized enough, however, that signals that are permanently excluded from conscious processing, as well as the facultatively-excluded signals from non-attended sensory stimuli, still have a strong impact on behaviour. Moreover, by influencing attentional mechanisms they can determine which of the stored memories or sensory signals will be transferred to the level of conscious processing. A hungry predator will search for traces of prey rather than mating partners, and so on.
One reason for the gated access of cognitive material to the level of awareness appears to be the limited capacity of the workspace of consciousness. Whether these limitations are due to the inability to attend to large numbers of items simultaneously, or whether they result from the restricted capacity of working memory, or even both, is subject to intense scientific investigation. The capacity of working memory is limited to about four to seven different items. The phenomenon of “change blindness”, which is the inability to detect local changes in two images presented in quick succession, demonstrates impressively our inability to attend to and consciously process all features of an image simultaneously. Because of these capacity constraints, conscious processing is in essence serial. Items are scrutinized and compared serially and therefore conscious processing is slow. Complex visual scenes are scanned serially and much of what we believe that we perceive simultaneously is actually reconstructed from memory. Which of the many signals finally reach the level of conscious awareness and can then be recalled depends on whether they are attended to, and this in turn is controlled either by external cues, such as the saliency of a stimulus, or by internal motifs, many of which we may actually not be aware of. And then it may occur that even an attentive, conscious search for content stored in declarative memory fails to raise it to the level of awareness. We are all familiar with the temporary inability to remember an episode or a name and have witnessed how a persisting subconscious search process suddenly lifts the content into the workspace of consciousness. It appears that we are not capable of controlling, at all times, which contents enter consciousness.
The differences between conscious and subconscious processes are further emphasized by evidence that the rules governing conscious deliberations and decisions most likely differ from those of subconscious processes. The former are based mainly on rational, logical, and syntactic rules, and the search for solutions is essentially based on serial computations. Arguments and facts are scrutinized one by one, and possible outcomes investigated. This strategy is suitable if variables are well defined, if sufficient time is available, if problems have a structure amenable to analytical treatment, and if precise solutions are required.
Subconscious mechanisms, by contrast, seem to rely more heavily on parallel processing, whereby a large number of variables enter into competition with one another. Then, a “winner takes all” algorithm leads to the stabilization of the activity pattern that is the most likely, given the initial conditions and the heuristics derived either from inborn routines or from past experience. The domains of subconscious processing are situations requiring very fast responses or conditions where large numbers of underdetermined variables have to be considered simultaneously, and weighed against variables that have no or only limited access to conscious processing—such as the wealth of implicit knowledge and heuristics, vague feelings, hidden motives, or drives. The outcome of such subconscious processes manifests itself either in immediate behavioural responses or in what is called “gut feelings”. And it is often not possible to indicate with rational argument why exactly one has responded in such a way and why one feels that something is wrong or right. In experimental settings one can even demonstrate that the rational arguments given for or against a particular response do not correspond to the “real” causes. For the solution of complex problems with numerous entangled variables it often turns out that the subconscious processes lead to better solutions than conscious deliberations—and this is thought to be because of the wealth of heuristics exploitable by subconscious processing. Given the large amount of information and implicit knowledge to which consciousness has no or only sporadic access, and given the crucial importance of subconscious heuristics for decision-making and guidance of behaviour, first identifying the structure of a problem and then deciding whether one should rely on conscious deliberations or listen to the voices of the subconscious appears to be a well adapted strategy.
However, because the two systems operate according to different principles, the solutions to a particular problem may not always agree. Most of the decisions that get us through daily life rely on subconscious processing and follow well-adapted heuristics. If these decision processes do not lead to immediate action, they may still influence subsequent behaviour by manifesting themselves as what we call “gut feelings”. One has no conscious recollection of the reasons that lead to these feelings, but one clearly experiences the reactions of one’s autonomous nervous system when the results of subconscious processes are in conflict with the outcome of conscious deliberations. In such situations one tends to say: “I decided according to all the rational arguments that I was aware of and took the best decision I could think of, but it somehow feels wrong.” The opposite situation is also possible, “I did what felt right to me, but if I think about it, it is absolutely crazy and irrational”. It is only when the two decision systems converge on the same solution one feels good, satisfied, and to some extent “free”.
After this brief excursion into the phenomenology of conscious and subconscious processes, intended to convey some connotations of consciousness, some of the most popular hypotheses about the constitution of consciousness in the brain will be reviewed.
One class of theories focuses on the philosophical implications of the hard problem of consciousness without attempting to provide detailed descriptions of putative neuronal mechanisms (Searle 1997; Metzinger 2000; Dennett 1992; Chalmers 2000). Solutions to the hard problem have also been sought through transcending current concepts of neuronal processes and incorporating theories borrowed from other scientific disciplines. The most prominent of these approaches assumes that phenomena unravelled by quantum physics also play a role in neuronal processes, and that they might be able to account for the emergence of consciousness from material interactions in the brain (Hameroff 2006; Penrose 1994). None of the predictions of these theories are at present amenable to experimental verification, because there is no evidence that quantum phenomena such as entanglement, superposition and collapse of wave functions, etc., play a role at the macroscopic level of neuronal network functions. Quantum effects do of course exist at the level of molecular and submolecular interactions, but it appears highly unlikely that they are relevant for the macroscopic functions of neurons responsible for information processing. Thus quantum theories of consciousness attempt to explain one poorly understood phenomenon with still unexplored and unproven mechanisms, and will therefore not be discussed further here.
Another class of theories pursues more modest goals and attempts to examine neuronal mechanisms potentially capable of supporting awareness of cognitive contents. Their aim is to define the neuronal mechanisms supporting the unitary character of awareness, its coherence in space and time, and the control of states that distinguish between conscious and unconscious processing (for review of relevant experimental findings see Melloni & Singer 2011; Aru et al. 2012a).
The most intuitively plausible solution for the unity of awareness is convergence of the results obtained in distributed processing areas to a singular structure at the top of the processing hierarchy. Theories derived from this intuition predict the activation of specific cortical areas when subjects are aware of stimuli. Consequently, these regions should remain inactive during unconscious processing of the same material. Likewise, lesions of these putative areas should abolish the ability to become aware of perceptual objects. So far, a region with such universal “observer functions” has not been identified, and this option is considered theoretically implausible by some (Dennett 1992). There is also little—if any—experimental evidence for such a scenario. If brain lesions abolish the functions of sensory areas or regions involved in the recall of memories, patients lose the ability to consciously experience the respective sensory contents or memories, but the ability to process other material consciously remains unaffected. Moreover, behavioural and brain imaging studies have shown that unconscious processing engages very much the same areas as conscious processing, including the frontal and prefrontal cortices (Lau & Passingham 2007; van Gaal et al. 2008). Thus there is no compelling evidence for specific areas supporting conscious processing. The are prominent examples of the selective elimination from conscious perception of those aspects of the stimulus material that are processed in specific regions without affecting awareness of other contents, such as syndromes of agnosia and blindsight (Cowey & Stoerig 1991), which result from selective lesions of sensory subsystems.
There are, however, systems and pathways in the brain whose destruction abolishes all conscious experience—but these cannot be considered to be the NCC. Rather, these systems adjust the narrow dynamic range within which the brain has to be kept in order to be operational and to perform the computations that ultimately give rise to awareness. These systems are addressed as modulatory systems; they originate mainly in deep structures of the brain and control global brain states via widely-diverging ascending projections.
Another class of theories favours the notion that the mechanisms supporting awareness of stimulation material are distributed and do not require anatomical convergence (Rodriguez et al. 1999; Metzinger 2000; Varela et al. 2001). Baars (1997) and Dehaene et al. (2006) propose that there is a workspace of consciousness whose neuronal correlate is a widely distributed network of neurons located in the superficial layers of the cortical mantel. As mentioned above, these neurons are reciprocally coupled through a dense network of cortico-cortical connections that have features of small-world networks. The proposal is that subjects become aware of signals if these are sufficiently salient to ignite coordinated activity within this workspace of consciousness. This is assumed to be the case for signals that either have high saliency because of the high physical energy of the stimuli or those that are made salient due to attentional selection.
Yet another, related proposal is that subjects become aware of contents, irrespective of whether they are triggered by sensory events or recalled by imagery from stored memories, if the distributed neurons coding these contents are organized into assemblies characterized by coherent, temporally-structured activity patterns. In this case, the critical state variable distinguishing conscious from non-conscious processing would be the spatial extent and the precision of coherence of temporally-structured neuronal responses (Rodriguez et al. 1999; Metzinger 2000; Varela et al. 2001).
In what follows, evidence will be reviewed in support of the latter hypothesis. However, before discussing this evidence we should briefly recall the reasons why temporal coherence should matter in neuronal processing.
Because of the small-world architecture of the cortical connectome, any neuron can communicate with any other neuron either directly or via just a few interposed nodes. Thus, efficient and highly flexible mechanisms are required, which permit selective routing of signals and assure that only the neurons that need to interact in order to accomplish a particular task effectively communicate with one another. Evidence from multisite invasive recordings and from non-invasive registration of global activity patterns with magneto-encephalography or functional magnetic resonance imaging indeed indicates that functional sub-networks are configured “on the fly” on the backbone of fixed anatomical connections in a task- and goal-dependent way. One mechanism that can accomplish such fast and selective association of neurons and gate neuronal interaction is the temporal coordination of oscillatory activity (Gray et al. 1989; Fries 2005). Since the discovery (Gray & Singer 1989) that spatially-distributed neurons in the primary visual cortex tend to engage in oscillatory responses in the beta and gamma frequency band when activated by appropriately configured contours, and that these oscillatory responses can synchronize over large distances within and across cortical areas and even hemispheres, numerous studies have confirmed that oscillations and their synchronization in different frequency bands are an ubiquitous phenomenon in the mammalian brain. The pacemakers of these oscillations are reciprocal interactions in local networks of inhibitory and excitatory neurons. The long-distance synchronization of this oscillatory activity appears to be achieved by several mechanisms operating in parallel: Long-range excitatory cortico-cortical connections, long-range inhibitory projections, and pathways ascending from nuclei in the thalamus and the basal forebrain (for a review of these see Uhlhaas et al. 2009). When neurons engage in oscillatory activity, they pass through alternating cycles of high and low excitability. At the peak of an oscillation cycle neurons are depolarized, highly susceptible to excitatory input, and capable of emitting action potentials. In the subsequent trough of the cycle, the membrane potential is hyperpolarized and membrane conductance is high because of strong GABAergic inhibition generated by the rhythmically-active inhibitory interneurons. During this phase, neurons are less susceptible to excitatory inputs, because excitatory postsynaptic potentials (EPSPs) are shunted and because the membrane potential is far from the threshold. Hence, neurons are unlikely to respond to pre-synaptic excitatory drive.
These periodic modulations of excitability can be exploited in order to gate communication among neurons. By adjusting oscillation frequency and phases of coupled neuronal populations, communication among those neurons can either be facilitated or blocked. To form a functional network of distributed neurons it suffices to coordinate their oscillatory activity in such a way that signals emitted by neurons of this network impinge on other members of the network at times when these are highly susceptible to input. One way to achieve this is to entrain the neurons that should be bound into a functional network to engage in oscillations of the same frequency, to synchronize these oscillations, and to adjust the phases such that neurons that ought to be able to communicate can communicate.
Evidence from multi-site recordings indicate that neurons are indeed bound together into sometimes widespread functional networks through synchronization of their oscillatory activity in a task-dependent way (Salazar et al. 2012; Buschman et al. 2012). This supports the hypothesis (Gray et al. 1989; Singer 1999; Fries 2005) that synchronization of oscillatory neuronal activity is a versatile mechanism for the temporary association of distributed neurons and the binding of their responses into functionally coherent assemblies—which as a whole represent a particular cognitive content. Such a dynamic binding mechanism appears to be an economical and highly flexible strategy for coping with the representations of a virtually unlimited variety of feature constellations characterizing perceptual objects. Taking the unified nature of conscious experience and the virtually infinite diversity of possible contents that can be represented, the formation of distributed representations, through response synchronization, offers itself as a mechanism that allows for the encoding of ever-changing constellations of contents in a unifying format.
Synchronization is also ideally suited to contribute to the selection of contents for access to consciousness. It enhances the saliency of signals by concentrating spike discharges into a narrow temporal window. This increases the coincidence of excitatory postsynaptic potentials (EPSPs) in target cells that receive input from synchronized cell groups. Because coincident EPSPs summate much more effectively than temporally dispersed EPSPs, synchronized inputs are particularly effective in driving post-synaptic target cells. It is thus not unexpected that entrainment of neuronal populations in synchronized gamma oscillation is used for attention-dependent selection of input configurations (Fries et al. 2001b; Fries 2009).
If activation patterns that subjects can become aware of are indeed characterized by globally coherent states of those cortical regions that process the contents actually appearing as unified, we expect these states of awareness to be associated with large-scale synchronization of neuronal activity. Candidate frequency bands are gamma and beta oscillations, as these have been shown to serve the temporal coordination of cortical networks. By contrast, if subjects are not aware of the presented stimulus material, processing should remain confined to smaller sub-networks, which operate in relative isolation and are not integrated into globally coherent states. In this case one should observe only local synchronization of more circumscribed neuronal populations (see Varela et al. 2001).
Finally, adjustments of oscillation frequency and phases fulfil the requirement that assemblies representing consciously processed contents need to be reconfigured at an extremely fast rate. The contents of which subjects are aware can apparently change at a rapid pace, at least four times a second, if one considers that this is the frequency with which the direction of gaze changes during the scanning of natural scenes. Thus, assemblies representing contents that are consciously perceived must be reconfigurable at similarly fast timescales. Evidence suggests that cortical networks operate in a regime of self-organized criticality close to the edge of chaos (Shew et al. 2009). Dynamical systems operating in this range can undergo very rapid state changes, which are characterized by shifts in oscillation frequencies, synchronization, and phase.
As discussed previously (Aru et al. 2012a) experiments designed to identify the neuronal correlates of consciousness are often fraught with ambiguities. The most frequently used strategy for the identification of NCC is contrastive analysis. One creates perceptual conditions in which targets are consciously perceived in only a subset of trials, while making sure that physical conditions are kept as constant as possible. This strategy implies that detection tasks have been designed, which operate close to the perceptual threshold. This can be achieved by reducing the physical energy of the stimuli or by masking them. While subjects are engaged in such detection tasks, neuronal responses are measured and then trials are sorted depending on whether the subjects did or did not perceive the stimulus. By subtracting the average responses obtained in the two conditions from one another, those neuronal responses that occur only in the condition of successful detection can be isolated, and these are then commonly interpreted as the neuronal correlate of conscious perception. This seemingly simple approach is not without ambiguity. Thus, noise fluctuations in afferent pathways are likely to lead to significant differences in the available sensory evidence, especially because experiments are performed at the perceptual threshold. Therefore, those aspects of neuronal responses that truly reflect NCC may be contaminated by signals resulting from noise fluctuations at processing stages preceding those actually mediating awareness. In addition, once subjects have become aware of stimuli, there are a number of subsequent processing steps that need not necessarily be linked to NCC. These comprise the covert verbalization of stimulus material, the engagement of working memory, the transfer of information into declarative memory, and perhaps also the preparation of covert motor responses. The distinction between these various confounding factors is difficult because all the processes are intimately related to each other. A detailed discussion of this problem is given in Aru et al. (2012a). One distinguishing feature could be the latency of the electrographic signatures of these various processing steps. Noise-dependent fluctuations in sensory evidence should be manifested early on; responses related to NCC proper should have some intermediate latency; and the consequences of having become aware of a stimulus should have the longest latencies. In order to use these latencies as distinguishing criteria, it is of course required that we estimate the precise latency at which the mechanisms leading to conscious perception are likely to be engaged. Assuming that the time required to prepare and execute simple motor responses is generally constant, the interval of interest can be constrained and has been proposed as somewhere between 180 and a few hundred milliseconds, depending on the sensory modality and the difficulty of the detection task. Attempts to use latency criteria for the elimination of confounds is of course restricted to electroencephalographic and magneto-encephalographic data, and cannot be applied to results obtained with functional magnetic resonance imaging because of the limited temporal resolution of this technique.
Another option for the reduction of confounds is to combine manipulations that influence the conscious perception of a stimulus through different mechanisms and to compare the electrographic responses between conditions (Aru et al. 2012b). We applied this strategy in investigations of patients with subdurally implanted recording electrodes located over the visual cortex. In one set of trials the visibility of stimuli, in this case faces, was manipulated by changing the sensory evidence of the stimulus material. In another set of trials visibility of the same stimuli was influenced by allowing the subjects to familiarize themselves with some of the stimuli. This also facilitated detectability, but now because of an expectancy-driven top-down process. The reasoning behind this was that neuronal responses reflecting NCC proper should be the same irrespective of whether stimuli are consciously perceived because of enhanced sensory evidence or because of top-down facilitation. As electrographic signature of interest we analyzed the neuronal activity in the gamma band. In a previous study (Fisch et al. 2009) had shown that category-specific gamma band responses in the visual cortex correlate with conscious perception. Conscious recognition leads to a phasic enhancement of the gamma-band response, thus supporting the notion that conscious perception arises locally within sensory cortices—which is in line with previous conclusions (Zeki 2001; Malach 2007). In our study we found that the performance and the reports of the subject were clearly modulated both by changing sensory evidence and by prior knowledge of the stimuli, as expected; but the gamma-band responses solely reflected the sensory evidence. This suggests that the differential activation of specific areas of the visual cortex, in our case mainly the fusiform face area, reflect processes that prepare access to conscious perception but are not its substrate proper.
Another frequently used paradigm in the search for NCC is interocular rivalry. If the two eyes are presented with stimuli that cannot be fused into one coherent percept, subjects perceive only one of the two stimuli at a time, and these percepts alternate. There are various ways to label the stimuli presented to the two eyes, to trace the responses related to their processing in the brain, and then to see which brain structures have to get involved in order to support conscious perception. Again, these studies have led to inconclusive results. Some claim that suppression of signals corresponding to the non-perceived stimulus occurs only at very high levels of visual processing, such as, for example, the temporal cortex, which is the highest stage of the ventral processing stream. The conclusion of these studies is that activation of this particular cortical network is a necessary prerequisite for conscious processing (Logothetis et al. 1996; Silver & Logothetis 2004). Others, by contrast, found diverging activity patterns already existing at the level of the thalamus and the primary visual cortex (Haynes et al. 2005; Fries et al. 1997). Recent correlations between the dynamics characterizing binocular rivalry and anatomical features of the primary visual cortex and the commissures linking the primary visual cortices of the two hemispheres provide compelling evidence that the rivalry phenomenon is based on processes occurring within V1 (Genc et al. 2014). However, none of these studies allows us to unambiguously locate the processes that lead to conscious perception. They only contribute to the identification of the earliest levels of processing, in which changes are detectable that correlate with conscious perception.
Interesting and of potential relevance for interpretations given in the next subsection is the observation that the access of sensory signals to conscious processing does not seem to be gated by modulation of the neurons’ discharge rate, but rather by changes of the synchronization of their activity—at least in the early stages of processing. What matters is the degree of synchronicity of oscillatory activity in the gamma frequency range. Signals conveyed by well-synchronized neuronal assemblies have access to conscious processing, while signals conveyed by similarly active but purely synchronized neurons fail to do so (Fries et al. 2001c). Of interest in this context is the observation that stimuli access conscious perception more easily if they are attended to, and that attention enhances synchronization of neuronal responses in the gamma frequency band in early visual areas (Fries et al. 2001b). Again, however, this local increase in synchrony is likely to simply enhance the saliency of the neuronal responses, facilitating their propagation across the cortical networks, and cannot per se be considered a neuronal correlate of consciousness.
The results described in what follows were obtained in a study where we presented words that could be perceived in some trials and not in others (by adjusting the luminance of masking stimuli), and simultaneously performed electroencephalographic (EEG) recordings (Melloni et al. 2007). Several measures were analyzed: Time-resolved power changes of local signals; the precision of phase synchronization across recording sites, and over a wide frequency range; and event-related potentials (ERPs). A brief burst of long-distance synchronization in the gamma-frequency range between occipital, parietal, and frontal sensors was the first event that distinguished seen from unseen words at about 180ms poststimulus. In contrast, local synchronization was similar between conditions. Interestingly, after this transient period of synchronization, several other measures differed between seen and unseen words: We observed an increase in amplitude of the P300 ERP for visible words, which most likely corresponds to the transfer of information to working memory. In addition, during the interval period—in which visible words had to be maintained in memory—we observed increases in frontal theta oscillations. Theta oscillations have been related to maintenance of items in short-term memory (Jensen & Tesche 2002; Schack et al. 2005).
To test whether the increase in long-distance synchronization relates to awareness or depth of processing, we further manipulated the depth of processing of invisible words. It has previously been shown that invisible words can be processed up to the semantic and motor level (Dehaene et al. 1998). In a subliminal semantic priming experiment we briefly presented words (which were thus invisible) that were either semantically related or unrelated, alongside a second, visible word on which subjects had to carry out a semantic classification task. Invisible words were processed up to semantic levels, as revealed by modulation of the reaction times, depending on the congruency between invisible and visible words: congruent pairs exhibited shorter reaction times than incongruent ones. We observed increases in power in the gamma frequency range for unseen but processed words. For visible words we additionally observed increases in long-distance synchronization in the gamma-frequency range (Melloni & Rodriguez 2007). Thus, local processing of stimuli is reflected in increases in gamma power, whereas long-distance synchronization seems to be related to awareness of stimuli. This suggests that conscious processing requires a particular dynamical state of the cortical network. The large-scale synchronization that we observed in our study could reflect the transfer of contents into awareness and/or their maintenance. We favour the first possibility, given the transient nature of the effect, and argue that the subsequent theta oscillations might support maintenance. It is conceivable that short periods of long-distance synchronization in the gamma band reflect the update of new contents, while the slower pace of theta oscillations might relate to sustained integration and maintenance of local results in the workspace of consciousness. The interplay between these two frequency bands might underlie the phenomenon of continuous but ever-changing conscious experience (see below).
More recently, Gaillard et al. (2009) have revisited the processing of visible and invisible words. In intracranial recordings in epileptic patients, they observed that invisible words elicited activity in multiple cortical areas, which quickly vanished after 300 ms. In contrast, visible words elicited sustained voltage changes, increases in power in the gamma band, as well as long-distance synchronization in the beta band and long-range Granger causality. In contrast to our study, Gaillard et al. observed a rather late (300–500 ms) rise of long-distance synchronization. However, it is important to note that in the study undertaken by Gaillard et al., phase-synchrony was analyzed for the most part over electrodes within a given cortical area, or at most between hemispheres. It is thus conceivable that earlier synchronization events passed undetected because of incomplete electrode coverage. Despite these restrictions, this study provides one of the most compelling pieces of evidence for a relation between long-distance synchronization and consciousness.
Some results of the experiments on binocular rivalry point in the same direction. Several studies have shown increased synchronization and phase locking of oscillatory responses to the stimulus that was consciously perceived, and controlled behaviour (Cosmelli et al. 2004; Doesburg et al. 2005; Fries et al. 1997; Srinivasan et al. 1999). Cosmelli et al. (2004) extend the findings obtained in human subjects by performing source-reconstruction and analyzing phase-synchrony in source space. These authors observed that perceptual dominance was accompanied by co-activation of occipital and frontal regions, including the anterior cingulate and medial frontal areas. Recently, Doesburg et al. (2009) have provided evidence for a relation between perceptual switches in binocular rivalry and theta- and gamma-band synchronization. Perceptual switches were related to increments in long-distance synchronization in the gamma band between several cortical areas (frontal and parietal) that repeated at the rate of theta oscillations. The authors suggest that transient gamma-band synchronization supports discrete moments of perceptual experience, while theta oscillations structure their succession in time, pacing the formation and dissolution of distributed neuronal assemblies. Thus, long-range gamma synchronization locked to on-going theta oscillations could serve to structure the flow of conscious experience, allowing for changes in content every few hundred milliseconds. Further research is required to clarify the exact relation between the two frequency bands, their respective role in the generation of percepts, and the pacing of changes in perception.
Another paradigm in consciousness research exploits the attentional blink phenomenon. When two stimuli are presented at short intervals among a set of distractors, subjects usually detect the first (S1) but miss the second (S2) when it is presented 200–500 ms after S1. Increases in long-range neuronal synchrony in the beta and gamma frequency ranges have been observed when the S2 is successfully detected (Gross et al. 2004; Nakatani et al. 2005). Furthermore, Gross et al. (2004) observed that successful detection of both S1 and S2 was related to increased long-distance synchronization in the beta range to both stimuli, and this enhanced synchrony was accompanied by higher de-synchronization in the inter-stimulus interval. Thus, de-synchronization might have facilitated the segregation of the two targets, allowing for identification of the second stimulus (also see Rodriguez et al. 1999). Source analysis revealed, as in the case of binocular rivalry, dynamical coordination between frontal, parietal, and temporal regions for detected targets (Gross et al. 2004).
In summary, studies of masking, binocular rivalry, and the attentional blink support the involvement of long-range synchronization in conscious perception. Recent investigations have suggested further that a nesting of different frequencies, in particular of theta and gamma oscillations, could play a role in pacing the flow of consciousness. Furthermore, the study of Gross et al. (2004) suggests that de-synchronization could serve to segregate representations when stimuli follow at short intervals. These results are encouraging and should motivate further search for relations between oscillatory activity in different frequency bands and consciousness, whereby attention should be focused not only on the formation of dynamically-configured networks but also on their dissolution.
Of the numerous proposals on NCC, those favouring temporal coherence as the mechanism that integrates widely distributed processes appear to be the least controversial. They account best for the apparent discrepancy between the unity of conscious experience and the distributed organization of the brain, because they allow for the dynamic integration of information generated in parallel by spatially segregated processing modules. Large-scale synchronization of oscillatory activity has been identified as a candidate mechanism for the flexible coupling of widely-distributed neuron populations, and hence as a likely NCC. This variable has the advantage that it can be measured relatively directly in humans who are able to give detailed descriptions about their conscious experience. However, oscillations and synchrony seem to be mechanisms that are as intimately and inseparably related to neuronal processing in general as the modulation of neuronal discharge rates. Thus, without further specification these phenomena cannot stand up as NCC—at least when we disregard the triviality that consciousness does not exist without them. We propose that the spatial scale and the precision and stability of neuronal synchrony might be taken as more specific indicators of whether the communication of information in the brain is accompanied by conscious experience or not. In this framework, conscious experience arises only if information that is widely distributed within or across subsystems is not only processed and passed on to executive structures but also bound together into a coherent, all-encompassing, non-local but distributed meta-representation. This interpretation is compatible with views that take consciousness to be the result of the dynamic interplay of brain subsystems; one that allows for a rapid and highly-flexible integration of information provided by the numerous distributed subsystems that operate in parallel. This view resembles a proposal from Sherrington, formulated in his book The Integrative Action of the Nervous System (Sherrington 1906): “[p]ure conjunction in time without necessarily cerebral conjunction in space lies at the root of the solution of the problem of the unity of mind.” The additive value of conscious processing would then be the possibility of establishing in a unified data format ever-changing relations between cognitive contents—irrespective of whether they are read out from memory or induced by sensory signals, and irrespective of the sensory modality providing the signals. By virtue of this dynamic definition of novel relations, non-local meta-representations of specific constellations could be established that have the status of cognitive objects. Just as with any other distributed representation of contents, these could then be stored as distributed engrams by use-dependant modification of synaptic connections, and thus influence future behaviour. Thus, conscious processing would differ from non-conscious processing because it allows for the versatile binding of the previously unbound into higher-order representations. And if so, “conscious” processing would be functionally relevant and not merely an epiphenomenon. Further arguments supporting the functional role of conscious processing are presented in the following section.