* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download Decoding visual consciousness from human
Sensory cue wikipedia , lookup
Executive functions wikipedia , lookup
Emotional lateralization wikipedia , lookup
Artificial general intelligence wikipedia , lookup
Activity-dependent plasticity wikipedia , lookup
Neural oscillation wikipedia , lookup
Affective neuroscience wikipedia , lookup
Neuropsychology wikipedia , lookup
Neurolinguistics wikipedia , lookup
Neuromarketing wikipedia , lookup
Neuroanatomy wikipedia , lookup
Neurophilosophy wikipedia , lookup
Optogenetics wikipedia , lookup
Consciousness wikipedia , lookup
Visual selective attention in dementia wikipedia , lookup
Brain Rules wikipedia , lookup
Nervous system network models wikipedia , lookup
Aging brain wikipedia , lookup
Dual consciousness wikipedia , lookup
Neural coding wikipedia , lookup
Cognitive neuroscience of music wikipedia , lookup
History of neuroimaging wikipedia , lookup
Functional magnetic resonance imaging wikipedia , lookup
Human brain wikipedia , lookup
Cognitive neuroscience wikipedia , lookup
Neuroplasticity wikipedia , lookup
Cortical cooling wikipedia , lookup
Philosophy of artificial intelligence wikipedia , lookup
Neural engineering wikipedia , lookup
Embodied cognitive science wikipedia , lookup
Development of the nervous system wikipedia , lookup
C1 and P1 (neuroscience) wikipedia , lookup
Binding problem wikipedia , lookup
Neuropsychopharmacology wikipedia , lookup
Hard problem of consciousness wikipedia , lookup
Neuroeconomics wikipedia , lookup
Holonomic brain theory wikipedia , lookup
Animal consciousness wikipedia , lookup
Feature detection (nervous system) wikipedia , lookup
Inferior temporal gyrus wikipedia , lookup
Artificial consciousness wikipedia , lookup
Neural binding wikipedia , lookup
Neuroesthetics wikipedia , lookup
Time perception wikipedia , lookup
TICS-762; No of Pages 9 Opinion Decoding visual consciousness from human brain signals John-Dylan Haynes1,2,3 1 Bernstein Center for Computational Neuroscience, Charité – Universitätsmedizin Berlin, 10115, Germany Graduate School of Mind and Brain, Humboldt-Universität zu Berlin, 10099, Germany 3 Max-Planck-Institute for Human Cognitive and Brain Sciences, Leipzig, 04103, Germany 2 Despite many years of research on the neural correlates of consciousness (NCCs), it is still unclear how the detailed contents of consciousness are represented in the human brain. It is often assumed that specific contents of consciousness are encoded in dedicated core NCCs – one for each different aspect of conscious experience. Now, the approach of multivariate decoding provides a novel framework for studying the relationship between consciousness and content-selective processing in more detail. This approach makes it possible to assess how conscious experience is encoded in the brain and how the encoding of sensory information is affected when it enters awareness. Core neural correlates of consciousness One distinguishing feature of visual consciousness (see Glossary) is its vivid, experiential quality. Take, for example, the famous sketch by Ernst Mach in his Analyse der Empfindungen [1] (Figure 1a). It shows Mach’s attempt to capture the first-person perspective of his visual experiences (specifically, the distribution of light and dark regions in his visual field) while he is looking out into his study. The traces of his eyebrow, nose and moustache enhance the first-person feeling for the observer, as if one were Mach himself, looking out of his left eye. This image immediately makes it clear that phenomenal consciousness is composed of several complex, structured contents that range from the fine-grained patterns of light and dark to meaningful individual objects. Several theoretical approaches have been proposed to unravel how the detailed contents of consciousness are realized in the human brain. Typically, these proposals entail a distinction between two types of conditions for consciousness [2–4]: first, ‘enabling’ or ‘background’ conditions, such as subcortical neural mechanisms of wakefulness, that are necessary to make an individual awake and conscious of anything (these conditions are unspecific in the sense that they are required for a broad range or even all conscious experiences to occur); and second, ‘contentspecific’ conditions that are necessary for a specific category of conscious experiences (such as brightness, colour or motion sensations). These specific conditions are thought to constitute a ‘core neural correlate of consciousness (NCC)’, a minimal set of neurons that shows a tight mapping or ‘direct correlation’ [2] with a category of Corresponding author: Haynes, J.-D. ([email protected]). experiences; it is conceivable that each different category has a different core NCC. Here, the view is adopted that on the one hand, a core NCC is necessary for a specific conscious experience (Box 1), and on the other hand, its activity patterns exhibit a consistent mapping to specific experiences that can be viewed as an ‘encoding’ of the experience in question. A key criterion for a core NCC is that it shows a maximally ‘tight’ and ‘direct’ correlation with the specific contents it realizes [2,3]. A frequent example is the correlation of motion experiences with processing in the human motion area MT/V5 [4]. Motion perception activates MT/V5 [5], the stimulation of MT/V5 influences motion judgements [6] and creates motion hallucinations [7], and the removal of MT/V5 yields motion blindness (also known as akinetopsia) [8]. This plausibly suggests that MT/V5 is the core NCC for conscious motion perception [4]. However, a tight or direct correlation seems to imply more than showing a dependence of motion perception on activity in MT/ V5. Direct correlation refers to the ability of different states of a core NCC to explain individual-specific subtle differences between various percepts. This requires showing the link at a more fine-grained, content-based level, as will be outlined below. Glossary Binocular rivalry: when conflicting stimuli are presented to the two eyes, conscious perception can alternate spontaneously between the input to the left and to the right eye [21,54,58]. Consciousness: the term ‘consciousness’ is used in different ways. This article follows the schema employed in Ref. [78]. In one sense, consciousness refers to the state of being awake and responsive as opposed to being asleep, in a coma or under anaesthesia. In a second sense, consciousness refers to the moment-to-moment contents of a person’s experience, which is the focus of the current article. ‘Awareness’ is used interchangeably with the second sense of ‘consciousness’ to denote conscious experience. Global workspace theory (GWS): a theory that postulates the neural process underlying conscious experience is a global distribution of information throughout the brain [74,75]. Lower bound of information: an estimate of the minimal amount of information encoded in a population signal regarding a specific content of consciousness. Microconsciousness: the theory that perceptual experience depends only on representation in specific sensory brain regions [59]. Multivariate decoding: the attempt to infer something about a person’s mind (such as the contents of their conscious experience) from distributed patterns of their brain activity. Multivariate (as opposed to univariate) indicates the simultaneous measurement of activity from many locations or neurons in the brain [10,11]. Neural correlate of consciousness (NCC): neural processes that correlate with many different contents of consciousness are considered to reflect ‘enabling’ or ‘background’ NCCs, such as wakefulness. Neural processes that encode specific contents of consciousness are termed ‘core’ NCCs [2–5]. 1364-6613/$ – see front matter ! 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.tics.2009.02.004 Available online xxxxxx 1 TICS-762; No of Pages 9 Opinion Trends in Cognitive Sciences Vol.xxx No.x Figure 1. Encoding of the contents of consciousness in a core NCC. (a) This classic sketch by Ernst Mach shows his first-person experience while he is looking out into his study. Experiences can vary along several dimensions (shades of brightness, orientations, textures and so on) and hierarchical levels (simple features, intermediate shapes and complex objects). (b) Each particular class of experiences is presumably encoded in a specific core NCC. For intensities, such NCCs presumably reflect univariate encoding by the level of activity in certain neurons (e.g. contrast in the spike rate of cells in V1). A different coding scheme is multivariate, in which each sensation is encoded by a different pattern of activity in a population of neurons. Such multivariate coding schemes can be either sparse, meaning that each different sensation is encoded by a single, specialized cell (a ‘cardinal’ or ‘grandmother’ cell), or distributed where the entire population of cells participates in encoding of each sensation [24]. (c) The mapping relationship observed between sensations and states of the core NCC has to fulfil certain requirements. It has to assign one neural state to each sensation (totality) and assign a different neural state to each sensation (injectivity). A violation of injectivity whereby two different sensations are mapped to the same state of the core NCC would mean that the sensation could not be decoded from the neural state in a lossless fashion. Please note that the mapping need not be single valued, meaning that a conscious sensation can be mapped to two different neural states and still be decodable. This allows the possibility of multiple realization of a sensation by different states of the core NCC (as when, for example, different microscopic constellations of spikes lead to the same average spike rate). Criteria for mapping conscious experiences onto neural states A different view is to think of a core NCC as a neural carrier for each particular category of experiences (e.g. brightness, colour and motion) in which these experiences are encoded. ‘Encoding’, here, simply means that there is a stable mapping between states of the core NCC and conscious experiences. It does not imply that there is a level at which the signals encoded in a neural carrier need to be ‘read out’ or ‘interpreted’ by some later level in the system. The reason for replacing ‘correlation’ with ‘encoding’ is that encoding provides a more generic and more powerful framework for identifying core NCCs. This provides a natural link to recent developments in experimental neuroscience based on multivariate decoding and pattern recognition [9–24]. This promising approach could yield a much tighter link between hypothetical NCCs and conscious experiences. Box 1. Necessary conditions of conscious experiences The question of which neural processes are necessary for awareness is elusive and will, thus, be briefly discussed here. Under normal conditions, an intact primary visual cortex seems to be a necessary condition for conscious experience because it is the main entry point for visual information into the cortical visual system, and damage to V1 leads to complete blindness in the corresponding region of the visual field [81] (but for an exception, see Ref. [82]). This could give the impression that activity in V1 is necessary for any kind of visual experience. However, V1 can be bypassed and vivid object experiences can be induced by directly stimulating the object-processing regions of the temporal lobe [83]. Hence, primary visual cortex is not strictly necessary for conscious experiences of objects and, thus, cannot be the core NCC for experiences regarding objects. This shows that it can be difficult to infer necessity from brain lesion data. To make this clearer, it can help to distinguish between strict and weak necessity. What is meant by strict necessity of a core NCC of an experience is that every possible way of achieving a particular experience requires activity in that core NCC (or, more formally: a strictly necessary condition for an experience is a condition that is a necessary part of every set of sufficient conditions for creating this experience; see, for example, Ref. [84]). A weakly necessary condition for a specific experience is required only for some ways 2 of achieving that experience. Therefore, activity in V1 is weakly necessary because it is a necessary part of some causal chains, including those that normally lead to object experiences whereby activity passes through V1. Unfortunately, many lesion studies do not enable one to distinguish whether a loss of awareness of a particular feature following a lesion to an area [8] means that this area is necessary in a strict sense or necessary in a weak sense for awareness of that feature (see, for example, Ref. [9] for an exception to Ref. [8]). This definition has to be distinguished from the notions of specific and unspecific conditions (equivalent to the ‘core neural basis’ and ‘total neural basis’ of conscious experiences in Refs [2–4]). Specific conditions are those that are required for a particular content of consciousness but not for others (such as activity in the fusiform colour area). Unspecific conditions are those that are required for many or even all contents of consciousness (such as activity in the brainstem). In the terminology of this paper, the ‘core neural basis’ is a strictly necessary condition that allows for lossless decoding of a specific type of experience. Please note that some authors also allow for the realization of a specific experience by multiple neural systems, in which case the necessity used here is dropped in favour of a minimal set of sufficient conditions (for details, see Refs [2,3]). TICS-762; No of Pages 9 Opinion Trends in Cognitive Sciences Vol.xxx No.x Figure 2. Encoding of intensity by signals in early visual cortex. (a) The perceived intensity of simple visual features is, presumably, encoded in the firing rate of cells in early visual cortex. The solid line shows a saturating model function that relates the perceived magnitude of contrast to the physical contrast of a stimulus. The monotonic relationship between physical and perceptual contrast is maintained in the responses of cells in primary visual cortex (dotted line, fMRI signals from V1 [26]; dashed line, average single cell in V1 [79]; solid line, model derived from behavioural psychophysics [80]). Thus, it is possible to decode the perceived intensity from signals in this NCC. The fMRI signal that reflects signals from the entire population of cells in V1 has the advantage of also reflecting the shape of the perceived contrast function and, thus, it can also explain magnitude relations between stimuli. (b) An important requirement for a core NCC is that it encodes a certain dimension of experience under various conditions (’invariance’). For example, when a central grating is surrounded by a larger region of the same orientation, the perceived intensity of the central grating is perceived to be reduced compared to when it is embedded in an orthogonal grating (right two gratings, effect enhanced for clarity). This enables one to test whether signals in V1 really indicate the same contrast level under different conditions. The magnetoencephalography (MEG) responses from primary visual cortex were different for stimuli of the same physical contrast that seemed to be different (two stimuli on the right), but they were the same when the physical contrast of the gratings was adjusted to make them seem to match (bottom left stimulus shows orthogonal matching stimulus M with reduced centre contrast, and bottom right shows parallel standard stimulus S). Thus, signals in V1 indeed reflect the perceived contrast of stimuli [30]. Importantly, it has been observed repeatedly that intensity encoding in V2 and V3 also closely matched that in V1 [26,28] and, thus, there are currently several candidate populations that could encode perceived intensity. A core NCC has to have the representational accuracy to encode a class of sensations. For this, it needs to fulfil specific mapping criteria (Figure 1b,c). These guarantee that each different experience can always be explained by a different neural state. Take as an example a hypothetical encoding of brightness and contrast experiences by activity in early visual cortex [25–30]. To be able to explain all brightness sensations by activity, say, in V1, this area needs to adopt a different activity pattern for each distinguishable level of brightness a person can experience. If V1 does not take on a different state for each experience, it does not have the representational accuracy to fully encode the different experiences and, thus, cannot be the core NCC for conscious brightness perception. Thus, the mapping has to assign one different and distinguishable neural state to each different sensation (‘totality’) and it has to be invertible, in the sense that no two sensations can be mapped to the same neural state (‘injectivity’; Figure 1c). The criterion of totality, for example, is violated in the case of MT/V5 because damage to MT/V5 mainly affects fast (but not slow) motion percepts [31]. Thus, MT/V5 cannot be the only core NCC in which all motion experiences are encoded. An example of a total mapping is illustrated in Figure 2a, in which each contrast experience is associated with a different state of a neural carrier in V1. The second criterion of invertibility of a mapping can be assessed by testing whether a particular dimension of conscious experience can be decoded in a lossless way from a parameter of neural activity [10,11,30] (Figure 1c). Importantly, the new approach of multivariate pattern recognition of brain signals enables one to test directly for such decodability [10,11]. This considerably extends previous studies that have investigated rudimentary content selectivity [32–35], such as the encoding of stimulus presence versus absence [34]. Representational units A class of experiences can be mapped either to a single property of neural activity (such as the activity level of single cells in early visual cortex [27]; Figure 1b, Figure 2) or to a property defined across large populations of cells in an area (as is the case for object perception, whereby cells specialized for different object features are distributed across a larger brain region, yielding a distributed representation; Figure 1b, Figure 3a) [36,37]. For this reason, the decoding approach needs to take the spatial response pattern across the entire population of neurons into account to reveal the information encoded in a brain region. It is even conceivable that a core NCC might be a parameter defined across multiple populations of neurons (e.g. a correlation pattern between two neural ensembles [38]). It could also turn out to be a property of a subpopulation of neurons in a region. It is possible that the activity of some neurons in an area might show a constant mapping to an experience but the remaining neurons fail to do so. Take, for example, the responses of primary visual cortex to unperceived stimulus features [20,33,39–41] – for example, that subjects cannot tell the eye of origin of a monocular stimulus [40]. This unconscious stimulus feature could be encoded in monocular subpopulations of primary visual cortex, whereas other conscious stimulus features (such as brightness) could be encoded in 3 TICS-762; No of Pages 9 Opinion Trends in Cognitive Sciences Vol.xxx No.x Figure 3. Distributed encoding of percept-based information. (a) In monkeys, complex object features are encoded in a columnar fashion in the inferior temporal cortex. This indicates a distributed, multivariate coding scheme for objects [37]. (b) Distributed spatial pattern responses obtained from the human object-processing region lateral occipital complex (LOC) using fMRI reflect the perceived similarity between different objects [47] and, thus, go beyond simple encoding by also explaining the topology of perceived relationships between different objects. This requires a relationship-preserving or ‘homeomorphic’ mapping between a set of experiences and the states of their core NCC. (c) Pattern-based decoding of information related to low-level features: the orientation of visual stimuli is encoded in a columnar pattern in primary visual cortex (from Ref. [19], reprinted with modifications as in Ref. [10] with permission). The left figure shows an optical imaging map of V1, in which each colour indicates a local predominance of cells of one particular orientation. The spatial resolution of fMRI (black grid) is not sufficient to resolve individual orientation columns. However, when subjects view gratings of different orientation, a reliable patterning of the fMRI responses from V1 is observed, such that different voxels respond best to different orientations (second from right) [19]. This can be explained as a result of small fluctuations in the density of cells specialized for different orientations in different fMRI voxels (second from left). Using multivariate decoding techniques, the orientation-related information encoded in these spatial patterns can be read out, making it possible to reconstruct the orientation of stimuli, despite the lack of spatial resolution of fMRI signals (right; for details, see Ref. [19]). binocular subpopulations. Also, it has to be ensured that the observed sources of decoded information are not physiologically epiphenomenal (e.g. subthreshold activity). In early visual areas, conscious versus unconscious perception is mainly reflected in blood-oxygenation-level dependent functional magnetic resonance imaging (BOLD fMRI) signals and local field potentials, particularly at lower frequencies, but, interestingly, not in spiking activity [42,43]. This shows that great care is needed when interpreting neural signals [44] and their potential implications for the NCC. Invariance across conditions An important further criterion is that the decoding needs to hold up under several different experimental conditions. For example, brightness and contrast can be probed under various contextual conditions (lighting, context, masking, and so on) [27,29,30]. Different physical stimuli leading to the same brightness sensations would need to be 4 invariantly mapped to the same states of the core NCC under all conditions [30] (Figure 2b). Lossless decoding versus correlation The ideal mapping criteria employed here are very strict and go beyond the simple correlation often measured between consciousness and brain activity. Most studies report only partial, rather than perfect, correlation (see Refs [32,33] for examples). This requires only that a part of the variance in neural signals is explained by consciousness. Such a lax criterion yields too many neural populations that correlate with consciousness and, thus, could all potentially be the core NCC where a specific content is encoded. By contrast, a lossless decoding of conscious perception from a corresponding brain signal would require a perfect correlation. Another problem is that correlation assumes a linear mapping between brain states and sensations, which is not necessarily required for perfect decoding (i.e. an increase in brightness must not TICS-762; No of Pages 9 Opinion necessarily be encoded in a linear or even monotonic increase in the spike rate of cells). Instead, there could be a more complicated or even arbitrary mapping that can be assessed using information-theoretic measures, such as nonlinear multivariate decoding and mutual information [10–13]. These measures can be used to reveal a constant mapping of one variable to another without requiring that the mapping be systematic. Encoding of similarity In an extension of the previous point, however, a systematic mapping between states of the core NCC and conscious experiences [2,45] would have additional explanatory power. It could explain how similarities or relationships between different sensations are encoded by similar relationships between states of the core NCC, as in the case of perceptual spaces [14,15,46,47] (Figure 3b). Some studies even show that it might be possible to provide a complete model for the encoding of a particular category of experiences in a core NCC [16,17]. Physical versus perceived features In most cases, our knowledge about a person’s sensations is based on psychophysical judgements about external physical stimuli [48]. However, it is important to clarify whether a particular brain area has information about the physical stimulus features (e.g. the light energy reflected by an object) or about a person’s conscious experiences (e.g. the perceived brightness of the surface). The maximum information about physical stimuli should be reconstructable from the retina, where graded receptor potentials closely follow physical properties of incoming signals. However, the perceived properties of objects are often different from the physical properties, as is the case in contextual interactions in brightness and colour perception [27,29,30,49]. For example, the encoding of chromatic signals in the retina and in V1 does not match the subject’s conscious colour perception that exhibits colour constancy across different illumination conditions [49]. Thus, activity patterns in the retina and V1 cannot explain colour perception because there is no constant mapping between states in these areas and individual colour percepts (see Ref. [50] for a similar argument). In contrast, a region in the temporal cortex of macaques contains cells that closely match the four non-reducible ‘unique hues’ of colour perception [51]. However, even if V1 does not encode colour sensations, it could still encode other, simple features of conscious experience, such as brightness or contrast sensations. Lossless decoding and bridge laws As mentioned previously, the quality with which a neural signal encodes a particular conscious experience can be probed by attempting to decode the contents of consciousness from that neural signal. As also mentioned previously, the term ‘decoding’ (as used here) refers to brute-force statistical techniques for testing whether a systematic mapping holds between states of a core NCC and a category of experiences. This is achieved by testing whether a classifier can learn to assign labels (i.e. sensations) to brain states correctly [10,11]. However, it does Trends in Cognitive Sciences Vol.xxx No.x not imply that brain signals are ‘interpreted’ at a semantic level or that a homunculus is required to decode the signals. Ideally, for a given class of sensations (e.g. brightness sensations), one would find only one particular parameter of neural activity (e.g. spike rate in V1) that allows for a lossless decoding. Then, sensations of this particular type could only be explained by this activity, and every time a person has a brightness sensation, one could argue that this state is directly encoded in the spike rate of V1 (however, see Ref. [2] for potential exceptions). Thus, the mapping can be used to establish ‘rules of correspondence’ or ‘bridge laws’ that enable one to link statements about the core NCC to statements about conscious percepts [52]. Currently, however, the lossless-decoding criterion has to be seen as an empirical ideal because of limitations in spatial and temporal resolution of non-invasive brain-imaging techniques and limitations in coverage of cell populations, brain regions and brain-activity parameters in many electrophysiological studies (but see Refs [42,43]). However, to an approximation, the brain parameter currently enabling the highest decoding accuracy can play the part of a candidate NCC for a specific sensation. Such decoding-based approaches can now also be performed using distributed ensembles of fMRI voxels [9–17,19–23] (Figure 3c), thus providing, for the first time, a noninvasive way of assessing the information encoded in various cortical areas. However, it is important to realize that such methods can only reveal a lower bound of information encoded in a brain region. The link between encoding and consciousness The ideal criteria outlined earlier provide an important contribution to the search for core NCCs (see supplementary material online for a full list of criteria and corresponding studies). They enable one to test whether a specific candidate NCC fulfils the necessary requirements to encode a specific class of sensations. Importantly, however, the criteria also allow one to address the question of what happens to a core NCC when the corresponding feature it encodes enters consciousness. At first, it might seem obvious that when a conscious percept fades in and out of consciousness (say, during binocular rivalry), the content would need to fade in and out of the corresponding core NCC (Figure 4a, second from left). In line with this, several invasive electrophysiological studies in humans and monkeys have shown that content-selective processes in the brain are modulated by consciousness of the corresponding contents [53–56], possibly even in an all-or-nothing manner [57]. For example, it has been demonstrated repeatedly that during binocular rivalry and flash suppression, content-selective cells in higher visual areas modulate their activity when their preferred object fades in and out of consciousness [53,54,58]. This is consistent with the microconsciousness theory that postulates that all that is required for representation of a content in consciousness is that it is encoded in a specialized perceptual processing module [59]. However, there are also several observations that complicate this straightforward equation of encoding and consciousness, as outlined below. These only become clear when one takes 5 TICS-762; No of Pages 9 Opinion Trends in Cognitive Sciences Vol.xxx No.x Figure 4. Encoding of information and conscious perception. (a) Different models can be used to explain why a stimulus fails to reach consciousness. When encoding of a stimulus and access or read out are both intact, a stimulus is consciously seen. A loss of conscious perception could be due to a failure of encoding of the stimulus in a core NCC, a failure of access or a failure of re-representation of a stimulus in a global workspace. (b) Attention. It is possible to decode which of two superimposed visual line images a person is currently covertly attending to from pattern signals in visual areas V1 to V4 [19]. (c) Consciousness. Even when line stimuli (T) are rendered invisible by rapidly exchanging them with a mask (M), their orientation can still be recovered from signals in primary visual cortex, indicating that V1 is not the core NCC of orientation sensations [20]. Interestingly, there is no evidence for information related to invisible orientation stimuli in higher visual areas beyond V1, indicating that V2 and V3 match conscious perception and, thus, one of these areas might contain the core NCCs of orientation sensations. (d) Left: the information about visual object stimuli that can be decoded from population signals in monkey temporal cortex is strongly reduced when a stimulus is rendered invisible by backward masking. Right: the information about a stimulus feature encoded in a brain region can be estimated by comparing the correlation between the responses to the repeated presentation of the same image (‘same category’) to the correlation between different images [9,23]. This can be used to compare the information encoded when objects are seen, as opposed to unseen. When human subjects view complex shapes, the information that can be encoded in V1 does not depend on whether they have successfully seen it or not (‘correct’ versus ‘incorrect’), whereas the information encoded in object-processing LOC does. This has two important implications for the relationship between information and awareness: first, conscious object perception might fail even though the simple features providing input to the object-processing network are fully intact; second, the information encoded in LOC closely matches trial-by-trial fluctuations in conscious perception. a more fine-grained, content-based view of the encoding of visual information in hypothetical core NCCs. First, it needs to be assessed whether a hypothesized core NCC that modulates with awareness has the representational accuracy to fully encode all experiences of the type in question. For example, when a face enters awareness during binocular rivalry, the activity in the fusiform face area (FFA) is increased [58], which could be taken to 6 indicate that encoding in the core NCC for faces is decisive for regulating whether a face is seen or not. However, it has been debated whether the face-related information encoded in FFA is sufficient to explain conscious perception of face identity, in contrast to a region more anterior in the temporal cortex [22,60,61]. Simply monitoring activity in specialized processing regions (such as FFA and parahippocampal place area [PPA]) allows one to track contents TICS-762; No of Pages 9 Opinion only in a coarse way and does not reveal whether such a content-selective region contains the core NCC encoding the fine-grained details of conscious experience (such as the faces of specific individuals). Second, consciousness is a multi-level phenomenon that spans from simple sensations of a distribution of brightness in the visual field to complex object perception. The fact that signals in higher but not lower visual areas match conscious perception of objects during rivalry has long been taken as evidence that consciousness occurs late in the visual system [2,33]. However, this neglects the multi-level nature of consciousness that includes not only the highlevel object category but also the fine-grained spatial layout of shades of brightness and colours of which an object is composed (Figure 1). Importantly, high-level object-processing regions do not encode information about these lowlevel features because they respond invariantly when the same object is defined by different features [62,63]. This invariance of high-level responses to low-level features is an important computational achievement of higher-order regions, but it implies that the information about simple contents of consciousness such as brightness and contrast is lost so that they have to be encoded in other neural populations, presumably in earlier brain regions. Third, it has been shown that cells in MT can signal the conscious perception of their preferred direction of motion under one experimental condition but fail to do so under a different experimental condition [55]. This means a simple equation between encoding in content-selective cells and consciousness of a corresponding feature cannot be true. Furthermore, content-specific brain regions that are modulated by consciousness can also be activated by unattended and even unconscious, invisible contents [64–66]. This would indicate that the encoding of signals in a contentselective region does not always lead to conscious experience of the corresponding content. Fourth, the encoding of sensory signals can be differentially affected by visibility in early and high-level visual areas. Using multivariate decoding of fMRI signals (Figure 3c), it has now been shown that V1 continues to encode information about the orientation of simple stimuli when the stimulus is unattended or even made invisible by masking [19,20] (Figure 4b,c). If primary visual cortex encodes information about orientation stimuli that fail to reach consciousness, the states of primary visual cortex do not provide the mapping needed to explain our simple sensations, as has been debated previously [32,33,39–41]. The findings regarding objects are different. Reliable information about conscious object percepts is encoded at the cortical site of high-level object recognition [9,14– 16,23,47,67] (Figure 4d). However, when an object is rendered invisible by masking or rivalry, the information encoded about its identity can be strongly diminished [67,68]. Similarly, trial-by-trial fluctuations in the visibility of objects are reflected in the information encoded in the human object-processing area [24] (Figure 4d). Thus, consciousness in the case of high-level object perception seems to strongly disrupt the encoding of objects in the corresponding core NCC, whereas in early visual areas, the information about low-level visual features can remain intact even though a stimulus fails to be seen. This Trends in Cognitive Sciences Vol.xxx No.x indicates that different levels of information can be differentially affected by changes in the level of visibility. Information about the simple constituent features of objects in early visual areas might remain unaffected by changes in their visibility, even when the neural encoding of the objects themselves is lost. Taken together, a closer look at the detailed encoding reveals that the representation in modality-specific regions is not sufficient to explain why perceptual information enters consciousness. This indicates that additional processes are required that regulate which contents gain access to consciousness [4,69] (Figure 4a, second from right). These mechanisms are believed to be closely linked to attention [70] but clearly involve more than attentional selection [71–73]. A hypothetical mechanism that regulates access to consciousness is described by the global workspace theory (GWS). The GWS postulates that encoding of information in content-specific and modality-specific brain regions is not sufficient and that information has to be globally distributed throughout the brain to reach consciousness [74,75]. In line with this, when a visual stimulus crosses the threshold to consciousness, fMRI signals are increased not only in content-selective visual brain areas but also in specific prefrontal brain regions (such as dorsolateral prefrontal cortex and medial prefrontal cortex) [76,77]. However, to test the hypothesis of global distribution of sensory information, one would need to investigate the degree to which the processes in prefrontal cortex (PFC) indeed receive such information when it reaches consciousness. Instead, the activity in PFC could reflect purely unspecific processes without content-specific encoding. For many processes that correlate with consciousness, it remains unclear the degree to which they are content specific. This makes it difficult to test the global workspace theory. This is a particular problem for fMRI studies, in which detailed information on content specificity of neural processes is only rarely available, particularly for supramodal cortical areas. As outlined above, the recent advances in decoding-based analyses of brain signals now enable one to directly probe the representational accuracy of brain signals recorded from different areas [10,11] and monitor how the encoding of information changes under varying levels of conscious access to visual information. This could help reveal whether the global distribution of information occurs in a graded manner, as hypothesized previously [35]. It would be particularly informative to attempt to decode visual information from activity in prefrontal brain regions and assess whether more information is available in PFC when a stimulus reaches consciousness. Importantly, such studies could also help distinguish between different readings of the GWS. In one variant, the core NCC is in modality-specific cortex and the global distribution of information is an ‘additional process’ that regulates whether a sensory representation enters awareness. In a different variant, one might assume that information is completely (rather than partially) re-represented in PFC when it reaches consciousness. This would indicate that the core NCC that fully spans a perceptual space for the currently relevant perceptual features is located in PFC, rather than in 7 TICS-762; No of Pages 9 Opinion modality-specific areas. These and similar issues require a decoding-based approach that assesses the degree to which the processes involved in conscious perception encode and re-distribute sensory information. Taken together, multivariate decoding provides a novel and powerful framework for identifying core NCCs and for investigating the link between neural encoding and consciousness. Further work is needed to re-examine how the decoding of information can inform models of consciousness. The key questions that now need to be addressed from an information-theoretic perspective are: Which neural core NCCs encode specific dimensions of conscious experience most veridically? To what degree do different manipulations of awareness (masking, rivalry and attention) affect the encoding of conscious experiences? The GWS needs to be further tested by assessing whether the neural activity found in prefrontal cortex when stimuli reach awareness contains sensory information. This could reveal a global distribution or possibly even a re-representation of information. Thus, studying the neural encoding of contents will be a key stage in unravelling the neural mechanisms of consciousness. Acknowledgements The author thanks Michael Pauen, Jochen Braun and Frederique de Vignemont for their valuable comments on the manuscript. This work was funded by the Max Planck Society, the German Research Foundation and the Bernstein Computational Neuroscience Program of the German Federal Ministry of Education and Research. Appendix A. Supplementary data Supplementary material associated with this article can be found at doi:10.1016/j.tics.2009.02.004. References 1 Mach, E. (1886) Beiträge zur Analyse der Empfindungen. Fischer 2 Chalmers, D. (2000) What is a neural correlate of consciousness? In Neural Correlates of Consciousness: Conceptual and Empirical Questions (Metzinger, T., ed.), pp. 17–40, MIT 3 Koch, C. (2004) The Quest for Consciousness: A Neurobiological Approach. Roberts 4 Block, N. (2007) Consciousness, accessibility and the mesh between psychology and neuroscience. Behav. Brain Sci. 30, 481–548 5 Tootell, R.B. et al. (1995) Functional analysis of human MT and related visual cortical areas using magnetic resonance imaging. J. Neurosci. 15, 3215–3230 6 Salzman, C.D. et al. (1992) Microstimulation in visual area MT: effects on direction discrimination performance. J. Neurosci. 12, 2331–2355 7 Pascual-Leone, A. and Walsh, V. (2001) Fast backprojections from the motion to the primary visual area necessary for visual awareness. Science 292, 510–512 8 Zeki, S. (1991) Cerebral akinetopsia (visual motion blindness). A review. Brain 114, 811–824 9 Haxby, J.V. et al. (2001) Distributed and overlapping representations of faces and objects in ventral temporal cortex. Science 293, 2425–2430 10 Haynes, J.D. and Rees, G. (2006) Decoding mental states from brain activity in humans. Nat. Rev. Neurosci. 7, 523–534 11 Norman, K.A. et al. (2006) Beyond mind-reading: multi-voxel pattern analysis of fMRI data. Trends Cogn. Sci. 10, 424–430 12 Muller, K.R. et al. (2001) An introduction to kernel-based learning algorithms. IEEE Trans. Neural Netw. 12, 181–201 13 Kriegeskorte, N. et al. (2006) Information-based functional brain mapping. Proc. Natl. Acad. Sci. U. S. A. 103, 3863–3868 14 O’Toole, A.J. et al. (2005) Partially distributed representations of objects and faces in ventral temporal cortex. J. Cogn. Neurosci. 17, 580–590 8 Trends in Cognitive Sciences Vol.xxx No.x 15 Haushofer, J. et al. (2008) Multivariate patterns in object-selective cortex dissociate perceptual and physical shape similarity. PLoS Biol. 6, e187 16 Kay, K.N. et al. (2008) Identifying natural images from human brain activity. Nature 452, 352–355 17 Thirion, B. et al. (2006) Inverse retinotopy: inferring the visual content of images from brain activation patterns. Neuroimage 33, 1104–1116 18 Quiroga, R.Q. et al. (2005) Invariant visual representation by single neurons in the human brain. Nature 435, 1102–1107 19 Kamitani, Y. and Tong, F. (2005) Decoding the visual and subjective contents of the human brain. Nat. Neurosci. 8, 679–685 20 Haynes, J.D. and Rees, G. (2005) Predicting the orientation of invisible stimuli from activity in human primary visual cortex. Nat. Neurosci. 8, 686–691 21 Haynes, J.D. and Rees, G. (2005) Predicting the stream of consciousness from activity in human visual cortex. Curr. Biol. 15, 1301–1307 22 Kriegeskorte, N. et al. (2007) Individual faces elicit distinct response patterns in human anterior temporal cortex. Proc. Natl. Acad. Sci. U. S. A. 104, 20600–20605 23 Williams, M.A. et al. (2007) Only some spatial patterns of fMRI response are read out in task performance. Nat. Neurosci. 10, 685–686 24 Quiroga, R.Q. et al. (2008) Sparse but not ‘grandmother-cell’ coding in the medial temporal lobe. Trends Cogn. Sci. 12, 87–91 25 Ress, D. and Heeger, D.J. (2003) Neuronal correlates of perception in early visual cortex. Nat. Neurosci. 6, 414–420 26 Boynton, G.M. et al. (1999) Neuronal basis of contrast discrimination. Vision Res. 39, 257–269 27 Rossi, A.F. et al. (1996) The representation of brightness in primary visual cortex. Science 273, 1104–1107 28 Haynes, J.D. et al. (2004) Responses of human visual cortex to uniform surfaces. Proc. Natl. Acad. Sci. U. S. A. 101, 4286–4291 29 Cornelissen, F.W. et al. (2006) No functional magnetic resonance imaging evidence for brightness and color filling-in in early human visual cortex. J. Neurosci. 26, 3634–3641 30 Haynes, J.D. et al. (2003) Neuromagnetic correlates of perceived contrast in primary visual cortex. J. Neurophysiol. 89, 2655–2666 31 Zihl, J. et al. (1983) Selective disturbance of movement vision after bilateral brain damage. Brain 106, 313–340 32 Tong, F. (2003) Primary visual cortex and visual awareness. Nat. Rev. Neurosci. 4, 219–229 33 Rees, G. et al. (2002) Neural correlates of consciousness in humans. Nat. Rev. Neurosci. 3, 261–270 34 Sergent, C. et al. (2005) Timing of the brain events underlying access to consciousness during the attentional blink. Nat. Neurosci. 8, 1391– 1400 35 Kouider, S. et al. (2007) Cerebral bases of subliminal and supraliminal priming during reading. Cereb. Cortex 17, 2019–2029 36 Bonhoeffer, T. and Grinvald, A. (1991) Iso-orientation domains in cat visual cortex are arranged in pinwheel-like patterns. Nature 353, 429– 431 37 Fujita, I. et al. (1992) Columns for visual features of objects in monkey inferotemporal cortex. Nature 360, 343–346 38 Engel, A.K. and Singer, W. (2001) Temporal binding and the neural correlates of sensory awareness. Trends Cogn. Sci. 5, 16–25 39 Gur, M. and Snodderly, D.M. (1997) A dissociation between brain activity and perception: chromatically opponent cortical neurons signal chromatic flicker that is not perceived. Vision Res. 37, 377–382 40 Blake, R. and Cormack, R. (1979) On utrocular discrimination. Percept. Psychophys. 26, 53–68 41 He, S. et al. (1996) Attentional resolution and the locus of visual awareness. Nature 383, 334–337 42 Maier, A. et al. (2008) Divergence of fMRI and neural signals in V1 during perceptual suppression in the awake monkey. Nat. Neurosci. 11, 1193–1200 43 Wilke, M. et al. (2006) Local field potential reflects perceptual suppression in monkey visual cortex. Proc. Natl. Acad. Sci. U. S. A. 103, 17507–17512 44 Logothetis, N.K. (2008) What we can do and what we cannot do with fMRI. Nature 453, 869–878 45 Müller, G.E. (1896) Zur Psychophysik der Gesichtsempfindungen. Zeitschrift für Psychologie und Physiologie der Sinnesorgane 10, 1–82 TICS-762; No of Pages 9 Opinion 46 Palmer, S.E. (1999) Color, consciousness and the isomorphism constraint. Behav. Brain Sci. 22, 923–943 47 Edelman, S. et al. (1998) Toward direct visualization of the internal shape representation space by fMRI. Psychobiology 26, 309–321 48 Fechner, G.T. (1860) Elemente der Psychophysik. Breitkopf und Härtel 49 Conway, B.R. and Livingstone, M.S. (2006) Spatial and temporal properties of cone signals in alert macaque primary visual cortex. J. Neurosci. 26, 10826–10846 50 Crick, F. and Koch, C. (1995) Are we aware of neural activity in primary visual cortex? Nature 375, 121–123 51 Stoughton, C.M. and Conway, B.R. (2008) Neural basis for unique hues. Curr. Biol. 18, R698–R699 52 Nagel, E. (1979) The Structure of Science. Hackett 53 Kreiman, G. et al. (2002) Single-neuron correlates of subjective vision in the human medial temporal lobe. Proc. Natl. Acad. Sci. U. S. A. 99, 8378–8383 54 Leopold, D.A. and Logothetis, N.K. (1999) Multistable phenomena: changing views in perception. Trends Cogn. Sci. 3, 254–264 55 Maier, A. et al. (2007) Context-dependent modulation of single neurons in primate visual cortex. Proc. Natl. Acad. Sci. U. S. A. 104, 5620–5625 56 Reddy, L. et al. (2006) A single-neuron correlate of change detection and change blindness in the human medial temporal lobe. Curr. Biol. 16, 2066–2072 57 Quiroga, R.Q. et al. (2008) Human single-neuron responses at the threshold of conscious recognition. Proc. Natl. Acad. Sci. U. S. A. 105, 3599–3604 58 Tong, F. et al. (1998) Binocular rivalry and visual awareness in human extrastriate cortex. Neuron 21, 753–759 59 Zeki, S. and Bartels, A. (1999) Towards a theory of visual consciousness. Conscious. Cogn. 8, 225–259 60 Leopold, D.A. et al. (2006) Norm-based face encoding by single neurons in the monkey inferotemporal cortex. Nature 442, 572–575 61 Tsao, D.Y. et al. (2006) A cortical region consisting entirely of faceselective cells. Science 311, 670–674 62 Sáry, G. et al. (1993) Cue-invariant shape selectivity of macaque inferior temporal neurons. Science 260, 995–997 63 Kourtzi, Z. and Kanwisher, N. (2001) Representation of perceived object shape by the human lateral occipital complex. Science 293, 1506–1509 64 Rees, G. et al. (2002) Neural correlates of conscious and unconscious vision in parietal extinction. Neurocase 8, 387–393 65 Moutoussis, K. and Zeki, S. (2002) The relationship between cortical activation and perception investigated with invisible stimuli. Proc. Natl. Acad. Sci. U. S. A. 99, 9527–9532 Trends in Cognitive Sciences Vol.xxx No.x 66 Fang, F. and He, S. (2005) Cortical responses to invisible objects in the human dorsal and ventral pathways. Nat. Neurosci. 8, 1380– 1385 67 Rolls, E.T. et al. (1999) The neurophysiology of backward visual masking: information analysis. J. Cogn. Neurosci. 11, 300–311 68 Sterzer, P. et al. (2008) Fine-scale activity patterns in high-level visual areas encode the category of invisible objects. J. Vis. 8, 1–12 69 Block, N. (1995) On a confusion about a function of consciousness. Behav. Brain Sci. 18, 227–287 70 Simons, D.J. and Rensink, R.A. (2005) Change blindness: past, present, and future. Trends Cogn. Sci. 9, 16–20 71 Braun, J. and Julesz, B. (1998) Withdrawing attention at little or no cost: detection and discrimination tasks. Percept. Psychophys. 60, 1–23 72 Lamme, V.A.F. (2003) Why visual attention and awareness are different. Trends Cogn. Sci. 7, 12–18 73 Koch, C. and Tsuchiya, N. (2007) Attention and consciousness: two distinct brain processes. Trends Cogn. Sci. 11, 16–22 74 Baars, B.J. (1997) In the Theater of Consciousness: The Workspace of Mind. Oxford University Press 75 Dehaene, S. and Naccache, L. (2001) Towards a cognitive neuroscience of consciousness: basic evidence and a workspace framework. Cognition 79, 1–37 76 Dehaene, S. et al. (2001) Cerebral mechanisms of word masking and unconscious repetition priming. Nat. Neurosci. 4, 752–758 77 Haynes, J.D. et al. (2005) Visibility reflects dynamic changes of effective connectivity between V1 and fusiform cortex. Neuron 46, 811–821 78 Zeman, A. (2007) Consciousness. In The Blackwell Companion to Consciousness (Velmans, M. and Schneider, S., eds), pp. 703–711, Blackwell 79 Albrecht, D.G. and Hamilton, D.B. (1982) Striate cortex of monkey and cat: contrast response function. J. Neurophysiol. 48, 217–237 80 Legge, G.E. and Foley, J.M. (1980) Contrast masking in human vision. J. Opt. Soc. Am. 70, 1458–1471 81 Holmes, G. (1918) Disturbances of vision by cerebral lesions. Br. J. Ophthalmol. 2, 353–384 82 Zeki, S. and Ffytche, D.H. (1998) The Riddoch syndrome: insights into the neurobiology of conscious vision. Brain 121, 25–45 83 Lee, H.W. et al. (2000) Mapping of functional organization in human visual cortex: electrical cortical stimulation. Neurology 54, 849–854 84 Mackie, J. (1974) The Cement of the Universe. A Study on Causation. Clarendon 9