Summary How is conscious experience related to material brain processes? A variety of…
- User
Summary How is conscious experience related to material brain processes? A variety of theories aiming to answer this age-old question have emerged from the recent surge in consciousness research, and some are now hotly debated. Although most researchers have so far focused on the development and validation of their preferred theory in relative isolation, this article, written by a group of scientists representing different theories, takes an alternative approach. Noting that various theories often try to explain different aspects or mechanistic levels of consciousness, we argue that the theories do not necessarily contradict each other. Instead, several of them may converge on fundamental neuronal mechanisms and be partly compatible and complementary, so that multiple theories can simultaneously contribute to our understanding. Here, we consider unifying, integration-oriented approaches that have so far been largely neglected, seeking to combine valuable elements from various theories. Introduction Comparing theories The mind-brain problem, i.e., how our conscious experience is related to material brain processes, has been debated by philosophers for centuries and remains one of the deepest unsolved problems in science.1,2,3 The last decades have seen a surge of theoretical and empirical consciousness research. A variety of neuroscientific theories of consciousness have been proposed and are hotly debated (e.g., Seth and Bayne4). Efforts to test competing theories against each other through “adversarial collaboration” have been undertaken, e.g., by consortia sponsored by the Templeton World Charity Foundation (e.g., Cogitate Consortium et al.5). Historically, most researchers have emphasized the development and validation of their preferred theoretical framework in isolation. This article, however, takes a different approach. It arises from researchers within the Human Brain Project (HBP; 2013–2023) working on topics related to consciousness and representing different theories. Based on recent empirical progress, we show that many aspects of the various theories of consciousness do not necessarily contradict each other, as sometimes claimed; instead, theories often try to explain different aspects of consciousness and tend to converge on fundamental neuronal mechanisms and processes. Here, we argue that several theories are at least partly compatible and complementary and we consider approaches toward convergence that have so far been largely neglected. As it is not feasible to cover all proposed, neuroscience-based theories of consciousness, we focus here on five that we think represent prominent and complementary perspectives: (1) global neuronal workspace theory (GNWT6), (2) integrated information theory (IIT7), (3) recurrent processing theory (RPT8), (4) predictive processing (PP9,10,11) and neurorepresentationalism (NREP12,13), and (5) dendritic integration theory (DIT14). These theories cover three of the four broad categories of consciousness theories outlined by Seth and Bayne4 and thus allow a broad discussion of how different theories may be reconciled and possibly integrated. Here, we use the term “theories” in a broad sense as there is no agreed-upon definition of this term (Table 1), and other terms (“hypothesis,” “model,” and “framework”) also lack stringent definitions. We focus on the content of these theories and ideas, rather than their status or maturity. Another criterion we used when choosing theories was that each should represent a cumulative body of neuroscientific research brought together under a definable framework of ideas. Term Definition or explanation References Comments/[synonyms or similar concepts] Conscious / consciousness (general and philosophical terms) “‘Conscious’ and ‘consciousness’ are umbrella terms that cover a wide variety of mental phenomena. … used with a diversity of meanings, … ‘conscious’ is heterogeneous in its range, being applied both to whole organisms... and to particular mental states and processes.” Stanford Encyclopedia of Philosophy Consciousness (C) (as used in this article) Despite the lack of any agreed upon definition of the term “consciousness,” we here use the term in a sense that is roughly synonymous with the term “experience,” including both PC and AC (see below) and thus including “subjective experience,” “inner experience.” Chalmers,1,15 Lamme,8 Pennartz,12 Koch et al.,16 and Sanders et al.17 The concept of consciousness that we use here includes both awake, alert states with their experiences, including perception of the external word, but also “inner experiences”: thoughts, feelings, dreams, imagery, hallucinations, etc. that may sometimes occur in unresponsive states (e.g., REM sleep, etc.). Phenomenal consciousness (PC) PC refers to those properties of experience that correspond to “what it is like” for a subject to have those experiences (qualia18). “These features are apparent to the subject from the inside, so tracking them arguably depends on one’s having the relevant experience.” Stanford Encyclopedia of Philosophy; Block19 That an item is part of PC does not imply that it is part of AC. Access consciousness (AC) AC refers to those properties of experience that are accessible for use in reasoning, report, and the control of action (i.e., in voluntary (“rational”) control of behavior). Stanford Encyclopedia of Philosophy; Block19 That an item is part of AC does not imply that it is part of PC. State of consciousness Overall conscious and unconscious states (related to brain states), e.g., wakefulness, dreaming, psychedelic states, dreamless sleep, minimally conscious or “vegetative” states, coma, general anesthesia. Stanford Encyclopedia of Philosophy [“Generic Consciousness: Stanford Encyclopedia of Philosophy”] Working memory (WM) A low-capacity memory system for temporary maintenance and manipulation of information in view of a task to be performed. It usually stores information for up to ∼10–15 s for use in reasoning, decision-making, and guiding of behaviors. Wang20 and Hempel and Oppenheim21 [Short-term memory (STM) is often used as a synonym of WM, but some researchers hold that STM is only short-term storage of information, whereas WM allows manipulation of the stored information in service of completing a task.] Sensory memory (SM) A large-capacity store, briefly storing (∼0.1–1 s) sensory data, separate for each sensory modality, providing a snapshot of the current, overall sensory input; automatic, outside cognitive control, weakly dependent on attention. Clark,9 Pennartz,13 and Carlson22 Includes iconic memory (for vision), echoic memory (for hearing), haptic memory (for touch stimuli). Explanandum The phenomenon to be explained by the theory; a sentence “describing the phenomenon to be explained” by the theory. Stanford Encyclopedia of Philosophy; Bastos et al.23 Scientific theory An explanation of an aspect of the natural world and universe that can be (or a fortiori that has been) repeatedly tested and corroborated in accordance with the scientific method, using accepted protocols of observation, measurement, and evaluation of results. Wikipedia (https://en.wikipedia.org/w/index.php?title=Scientific_theory&oldid=1183411663) There is no universally accepted definition of “scientific theory.” Similar definitions are, e.g., given by Encyclopedia Britannica (https://www.britannica.com/science/scientific-theory), Stanford Encyclopedia of Philosophy (https://plato.stanford.edu/entries/structure-scientific-theories/). National Academy of Sciences USA,24 and American Association for the Advancement of Science (https://www.aaas.org/sites/ default/files/0219boardstatement.pdf). Table 1 Terms and concepts
Open table in a new tabBy examining how these five theories compare and may be partly reconciled, we aim to start an integration process that could also incorporate valuable elements from other theories in the future. However, some current theories may of course turn out to be entirely or largely erroneous or sterile and will be weeded out by future selection, as has happened repeatedly in the history of science. Trying to overcome long-standing, seemingly unresolvable philosophical and theoretical debates, we propose a different approach, focusing on empirically based notions of integration, complexity, representation, and recurrent processing at different levels, which appear to be shared features of these theories. Following Ned Block,19 we distinguish mechanisms of phenomenal consciousness (PC; immediate, subjective experience, e.g., of seeing a blue sky) from those of access consciousness (AC) and reportability. We discuss how these distinctions are related to proposed empirical measures of consciousness (Boxes 1 and 2) and how they relate to attention, expectations, planned behavior, and different forms of memory.8,12,25,26,27,28 We also distinguish between levels of explanation from low-level, mechanistic features to high-level, system-wide properties. Box 1 Methods for assessing states and levels of consciousness Identifying brain-based measures to reliably index the presence and the absence of consciousness in humans is the focus of a substantial body of literature, a fundamental concern for clinicians, and arguably an important first step for understanding the relationship between consciousness and its physical substrate. Practically, brain-based measures are mandated when the behavioral assessment of consciousness is unreliable, such as in subjects who are paralyzed during general anesthesia or are unresponsive because of severe brain injuries. A first class of markers involves by-passing motor behavior and directly measuring brain responses to sensory stimuli, such as verbal commands or complex sequences of sounds. A prominent example of this class of markers is the P3b elicited by global violations of expectation in auditory sequences.29 According to GNWT, the P3b indexes the broadcasting (following ignition) of sensory information to frontal cortical processors and has been used to detect the presence of consciousness in brain-injured patients. Although the P3b provides a specific marker (i.e., it is present only in subjects who are aware of the stimulus), it is often absent in conscious brain-injured patients and conscious healthy subjects.30,31 Such low sensitivity of the P3b is an empirical reminder of the potential dissociation between the neuronal correlates of AC (indexing connectedness to the environment and/or engagement of high-level cognitive functions) and those of PC. A second class of markers focuses directly on brain activity, without necessarily relying on sensory processing and cognitive functions. A recent empirical convergence has emerged on measures of integration and differentiation holding the promise of a more accurate detection of consciousness.32 Notably, this convergence involved many independent researchers endorsing disparate theoretical frameworks and employing different techniques across various conditions such as sleep, dreaming, anesthesia, epilepsy, hallucinatory states, and coma. In general, the prospect of reliably detecting consciousness independently of connectedness to the external environment, behavioral performance, cognitive, and memory abilities has clinical implications and fosters the search for the neuronal mechanisms of PC. Complexity measures rely either on the characterization of spontaneous brain dynamics (through the analysis of functional connectivity networks and graph theory metrics)33,34 or on the quantification of the patterns of interactions evoked by direct cortical perturbations.35 The latter approach, instantiated by the PCI, shows high sensitivity in challenging clinical conditions and points to a mechanistic link between brain complexity and feedback interactions. Indeed, across various physiological and pathological conditions and experimental models, decreases and increases of PCI during loss and recovery of consciousness are invariably associated with the obliteration and resurgence of recurrent waves of activations.36,37,38,39 Box 2 Methods for assessing contents of consciousness Assessing the contents of consciousness empirically at a fine-grained sub-categorical level is challenging because recording sites representing specific features need to be identified. A useful approach has been proposed by Koch et al.16 when searching for content-specific neural correlates of consciousness (NCC): (1) a mapping step in which the response properties of recording sites (e.g., neurons in neuronal recording studies or voxels in fMRI studies) are determined in a participant using appropriate feature-specific stimuli; (2) application of an experimental paradigm tapping on consciousness such as a multistable figure or binocular rivalry where the physical stimulus properties remain constant, while the contents of consciousness alternate between different interpretations of the stimulus. If the experienced (reported) alternating content of consciousness is reflected in dynamic activity changes in the features mapped in step 1, the respective recording sites qualify as potential NCCs; (3) experimentally try to rule out known confounds, for instance, that neural activity covarying with conscious perception might instead reflect prerequisites or consequences of consciousness.40 Thus, it must be kept in mind that this approach is correlational, not causal, and that attention and memory effects may confound the interpretation of results. This approach has been successfully applied in macaque monkeys as well as in humans using electrophysiological recordings. Human intracerebral (stereo-EEG) recordings led to the identification of tonic activity patterns,41 i.e., prolonged low-amplitude gamma-band responses to sensory stimulation, likely corresponding to the intracranial correlate of previously reported late EEG components reflecting perceptual awareness.42,43 Somatosensory stimulation induces tonic activity in the bilateral parietal operculum and posterior insular cortices.44,45 In addition, visuo-tactile stimulation enhances specifically these tonic components,46 paralleling the notion that visuotactile stimulation considerably improves detection in tactile-extinction patients.47 The approach could be extended to human fMRI measurements with a sufficiently high spatial resolution to differentiate feature-specific responses at the level of cortical columns. In a study by Schneider et al.,48 columnar clusters in the human middle temporal complex (hMT+) that selectively responded to horizontal vs. vertically moving stimuli were localized in the mapping stage. In the main experiment, an ambiguous apparent motion stimulus (motion quartet) was presented that could be perceived as either horizontally or vertically moving. By tracking activity in the pre-mapped horizontal/vertical columnar clusters, it was possible to predict which of the two possible interpretations of the bistable motion stimulus (horizontal or vertical motion) was perceived by the participant at alternating periods in time. The results of the motion quartet study support the idea that hMT+ is (part of) the content-specific NCC for direction-of-motion, in line with the causal role of the middle temporal area (MT) in motion vision elucidated by Salzman and Newsome in non-human primates.49 More generally, mapping out cortical (columnar) functional clusters and measuring their responses using sub-millimeter ultra-high field fMRI during ambiguous sensory stimulation fulfills the criteria for the search of content-specific NCC outlined by Chalmers15 and Koch et al.16 Concepts and explananda When comparing theories, it is essential to first clarify the key concepts, define the terms used, and identify which phenomena (explananda) the theories aim to explain. A point that is immediately highlighted by our comparative approach is that different theories tend to focus on distinct aspects or subtypes of “consciousness,” which may cause conceptual confusion and miscommunication. Opposing theories may promote different experimental approaches that are not directly comparable, or different interpretations of data. For example, GNWT has focused on AC19 (Table 1), primarily trying to explain how conscious content can be accessed and used by the subject, and on conscious vs. unconscious “information processing” (as opposed to phenomenal experience), typically studying responses to a particular sensory input, such as seeing a face. By contrast, integrated information theory (IIT), recurrent processing theory (RPT), and predictive processing (PP)/NREP have primarily focused on “PC,”19 (i.e., what “it is like” to experience something18), irrespective of cognitive functions and whether it can be reported or not.50,51 Thus, in contrast to GNWT, the other theories (IIT, RPT, and PP/NREP; for dendritic integration theory (DIT), see below) all primarily try to address the “Hard Problem,1” i.e., to explain how it is possible that PC and “qualia” can arise in a physical system like the brain, and why these subjective phenomena disappear during dreamless sleep or anesthesia (but see Sevenius Nilsen et al.52). This difference partly comes from a different philosophical view: contrary to IIT, PP/NREP, or RPT, the proponents of GNWT typically do not regard PC or qualia as a real phenomenon, distinct from AC, but have argued that as long as their theory explains the observable facts (e.g., reports like: “now I see the face!”), there is nothing left to explain. For PP/NREP, IIT, and RPT, PC is the primary explanandum, regarding access to conscious content as an accessory matter, relying on executive processes that apply conscious content to motor or cognitive operations.7,12,13,27,53 DIT has suggested mechanisms for both AC and PC.54 The original versions of GNWT vs. IIT also differed in another way: GNWT focused mostly on the processes of selection and use of contents of consciousness (AC), whereas IIT focused more on the fundamental nature and qualities of PC (in terms of information integration and interactions), and the state and level of PC.55 However, now both theories increasingly try to explain processes, contents, and states/levels,28,16 a tendency shared with RPT, PP/NREP, and DIT.12,14,56 There is an increasing agreement that any complete theory should explain both contents and states/levels of consciousness.57 Finally, DIT is a recently proposed theory, mainly operating at the cell and circuit levels, and addresses why mammalian consciousness depends on both cortico-cortical and thalamo-cortical loops. It hypothesizes that both AC and PC involve simultaneous completion of these loops due to dendritic processing.14,54 In this review, we will first briefly present each of the five theories one by one, then discuss how they compare with each other, using the brain’s hierarchical organization as a scaffold, and next consider if and how different theories may be reconciled and perhaps even integrated. Theories of consciousness Global neuronal workspace theory The global workspace theory (GWT)58 describes the processing steps that give rise to conscious experience as a dynamic interaction between local processors and a global workspace that integrates information and broadcasts it to other processors for further processing. The global workspace thus acts as backbone through which information about different sensory and cognitive processes is made available to multiple brain regions. It is the broadcasted information that constitutes the content of consciousness. A neurobiology-inspired version of the GWT was later proposed as the GNWT.6,28,59,60 It takes the same functional concept of a global workspace but links it to specific brain networks and neuronal characteristics. Here, the local processors comprise modular cortical areas with functional specializations for perception, action, memory, etc. Their activity is triggered by sensory stimuli, and they can perform calculations in isolation from each other. Modules with such characteristics are, for instance, found in early sensory cortex, where different features are processed in specialized cortical regions.61,62,63 The global workspace is thought to involve long-range connections (cortico-cortical and thalamo-cortical) that connect a broad network of brain regions, particularly in the prefrontal cortex, posterior parietal cortex, and associated areas. These distributed areas can co-select one mental object so that all the representations of relevant features of that object are enhanced and bound in perception. The formation of coherent object representations corresponds to object-based attention in perceptual psychology. In this scheme, the thalamus and other subcortical structures such as the basal ganglia could also serve to amplify and distribute the information across various cortical regions and integrate neuronal activity across large brain regions. The cortical areas are believed to select and integrate information to enable the brain to focus and carry out complex cognitive tasks. Finally, the moment a local neuronal assembly reaches the activity threshold to enter the global workspace and engage in subsequent broadcasting is called “ignition.” Ignition is a central concept in GNWT that has been linked to recurrent processing, for instance, in frontoparietal cortex.64 It has gathered support in recent years,28 for instance, from human neurophysiological and neuroimaging studies that show a distinct divergence in brain activity around 200 to 300 milliseconds after stimulus onset that depends on conscious perception65,66 and has been observed across different sensory modalities and paradigms. Early perceptual processing in the first 200 milliseconds can be similar or even identical between conscious (i.e., reported) and unconscious (i.e., not reported) trials,67,68,69 but conscious perception is associated with late events, such as the N2/P3b/P300 component of scalp-event-related potentials in humans.6,70,71,72 The timing of conscious access can furthermore be delayed under certain conditions, such as inattention or dual-task situations where the global workspace is already occupied by another conscious object.67,73 Concept cells in the anterior temporal lobe of humans and prefrontal neurons in non-human primates show stronger and longer-lasting firing rates when corresponding stimuli are consciously perceived.74,75,76,77 In monkeys, signal propagation to the frontal cortex furthermore leads to ignition and behavioral reports.64 Some of these experimental results are debated,12,16,78 often on the basis that perceptual reports may confound the later frontal activity.42 However, results from recent no-report paradigms in monkeys may potentially mitigate these criticisms.79,80 Recurrent circuits within or across brain areas support persistent activity and ignition-like processes.20 Long-range cortical projections, crucial for the global workspace, originate in cortical layers II/III and V/VI where feedback effects are also most pronounced.81,82,83,84 Recent evidence suggests that reverberatory loops for persistent neuronal activity involve both cortico-cortical interactions and subcortical regions like the thalamus and cerebellar nuclei,85,86 which is in line with a broad network of long-range connections supporting a global workspace.87 Integrated information theory With its emphasis on the “phenomenal structure” of experience, IIT uses introspection to characterize phenomenal properties and then deduces these in terms of corresponding physical properties. This physical explanation can then be empirically validated or falsified according to the standard methods of measuring behavioral, functional, and neural correlates.50 The seed of this consciousness-first approach can be already found in the early paper “Consciousness and Complexity” published in 1998.88 Here, the authors start from introspection and the premise that each conscious experience is at once integrated (unitary) and highly informative (each conscious state differs in a specific way from a huge number of alternative states). They then postulate that the coexistence of neuronal integration and differentiation is a key physical property of conscious systems, operationalizing this claim as measurable neuronal properties: (1) Integration: a group of neurons can contribute to conscious experience only if it is part of a functional cluster that achieves high integration through reentrant interactions in hundreds of milliseconds. (2) Differentiation: to sustain conscious experience, it is essential that this functional cluster must be capable of a large repertoire of different neural states, associated with high values of complexity. This principle, relating consciousness to the complexity of neuronal interactions, is at the core of the early formulation of IIT,89,90 it has inspired a version of “weak IIT,91” and led to testable explanations and predictions: (1) Some regions of the thalamo-cortical system endowed with appropriate anatomical requirements (high density of lateral connections and large numbers of reentrant circuits) support conscious experience, whereas other structures with modular or feedforward architectures, such as the cerebellum, do not. (2) Synchronous or widespread neuronal activation is not necessarily associated with conscious states unless they are characterized by high levels of differentiation or complexity. (3) Given that anatomical requirements are preserved, functional alterations affecting reentry, integration, and differentiation will result in the loss of complexity and consciousness in conditions such as sleep, general anesthesia, and generalized seizures. (4) High-complexity and conscious experience can be supported by intrinsic brain interactions even in the absence of sensory inputs, motor outputs, and executive functions. (5) The presence vs. absence of consciousness can be reliably assessed by using direct cortical perturbations and recordings to measure the complexity of causal interactions within the brain.92 These early predictions have been extensively assessed over the last 20 years across different conscious states (wakefulness, sleep, dreaming, psychedelic states, epilepsy, and disorders of consciousness) using various brain imaging modalities and have shown promise of clinical application (for a review, see Sarasso et al.32; see also Box 1). The full formulation of IIT93,94 culminating in IIT 4.095 considers five phenomenal properties of consciousness (intrinsicality, information, integration, exclusion, and composition) and then formulates them as physical properties. Through its mathematical formalism IIT aims to account not only for the presence vs. absence of consciousness but also for its quality and contents. Specifically, IIT posits an explanatory identity between each experience—a particular “phenomenal structure”—and the corresponding “cause-effect structure” specified by the physical substrate.50 IIT holds that a cause-effect structure captures in full the causal powers of a system in a particular state—say, a network of neurons, some active and some inactive. Testing this framework requires combining systematic perturbations, observations, and partitions to assess precisely how subsets of elements make a difference to and take a difference from one another (Figure 2). Figure 1 Global neuronal workspace theory (GNWT) Show full captionFigure viewer Figure 2 Integrated information theory (IIT) Show full captionFigure viewer Although fully unfolding the cause-effect structure of even simple physical substrates is computationally daunting, IIT makes interesting predictions that are amenable to empirical tests. For example, IIT posits that (1) hierarchically organized grid-like cortical structures with lateral and recurrent convergent-divergent connections, such as those found in posterior cortex, are optimally suited for integrating information and thus correspond to the neural substrate of consciousness,96 and (2) that phenomenal space—the feeling of extendedness—can be accounted for by the “cause-effect structure” specified by the grid-like structure of cortical areas.97 (3) Within these grids, also inactive neurons contribute to the cause-effect structure and thus to experience, just like active neurons. (4) Finally, modular or feedforward architectures, such as deep learning networks, would not be conscious even if their observable performance should equal that of a conscious human.94 Recurrent processing theory The core idea of the RPT is that recurrent (feedback, reentrant) processing (RP) is essential and perhaps even sufficient for conscious experience. Victor Lamme developed RPT based on experiments on primate vision,98,99,100 leading to a general theory of visual consciousness.27,101,102 A key element of RPT is that it regards attention and consciousness as separate (orthogonal) processes (Figure 3, top, horizontal, and left, vertical arrows).8 This distinguishes RPT from many other theories, including GNWT, PP, and the attention schema theory (AST103), but not from NREP. Figure 3 Recurrent processing theory (RPT) of visual consciousness Show full captionFigure viewer RPT outlines four stages of visual processing (Figures 3A–3D): two are unconscious (stages 1 and 2) and two others (stages 3 and 4) are regarded as conscious, but different (phenomenal vs. AC; PC vs. AC).19 RTP also relates PC and AC to attention and different forms of memory (Figures 3C and 3D), thus pioneering attempts to reconcile RPT and other theories.8,25,105 Stage 1 Local (“superficial”) feedforward processing: a fast feedforward sweep (FFS, ∼100–200ms) of activity carries information from primary to higher visual cortices, where image features are extracted. However, if not boosted by attention, processing stays local, failing to proceed to stage 2, yielding only “subliminal” processing106 without PC or AC27,102,105 (Figure 3A). However, “non-feedforward” (recurrent/lateral/horizontal) processing can be elicited by the FFS at all cortical levels, thus triggering stages 3 and 4. Stage 2 More global (“deep”) feedforward processing occurs when the stimulus is attended. If so, the boosted FFS propagates up to motor and prefrontal areas, leading to rapid (∼200 ms) identification of the main content. This can trigger immediate, unconscious (“automatic”) behaviors, such as hitting a tennis ball (Figure 3B),27,102 while the subject is unaware of both stimulus and action. Stage 3 Local (“superficial”) RP is an intermediate processing stage where unattended, unmasked stimuli, mainly in the ventral visual stream,107 cause phenomenal conscious experience (PC19). However, without attention, the recurrent processing is limited and does not provide interactions with prefrontal and motor areas. The limited recurrence implies that sensory stimuli do not enter working memory and further cognitive processes that are needed for reporting. Hence, there is PC without AC (Figure 3C). Stage 4 Given sufficient time, attended stimuli lead to deep, global recurrent processing (Figure 3D),27,102,105 reaching higher levels, including prefrontal, executive areas, providing contents accessible to the brain’s executive systems (AC in addition to PC), for use in, e.g., working memory, reports, and other behavioral responses, i.e., “ignition” in GNWT (see above). Within RPT, attention is crucial for determining which stimuli cause global (stage 4) RP and AC, vs. local (stage 3) RP and PC. Only with attention does the subject “know” and can use and report its experience (AC). RPT also claims that different forms of short-term memory map onto PC and AC. Whereas global RP engages robust, low-capacity working memory (WM), local RP engages large-capacity, very brief, sensory (iconic), and fragile memory (SM, FM; 0.1-few seconds) in sensory cortices, which is easily overwritten by other stimuli.105 This explains, e.g., inattentional and change blindness, and why only a few visual items (held in WM) are reported after brief presentations with no cue.108 Being developed by experimentalists, RPT is firmly rooted in empirical observations, unlike more philosophical theories. Numerous experimental results support RPT, e.g., from masking, change blindness, inattention, attentional blink, neglect, and extinction,8,101,102,105 but do not exclude other theories. An early (<200 ms) event-related potential (ERP) component (visual awareness negativity; VAN) seems to match (in timing and topology) local RP (Stage 3104) and may reflect PC (Figure 3C). The late (∼300 ms) positivity (LP) ERP wave (P300/P3b29) shows characteristics of stage 4 and may thus reflect AC (Figure 3D),102,105 in agreement with GNWT.28 Similar early ERP waves are found also for other sensory modalities, suggesting a general perceptual awareness negativity (PAN) reflecting sensory PC.43 A challenge for RPT is whether RP “is sufficient for conscious experience.8,102” Although several empirical challenges that may seem to falsify RPT are resolved by distinguishing PC from AC,8,19 the core problem remains: RP seems far more ubiquitous than consciousness, suggesting that RP alone is not sufficient for consciousness.102 Indeed, RP is not completely lost in apparently unconscious states (dreamless sleep, anesthesia, coma;102; but see Sevenius Nilsen et al.52). Lamme suggests that the “missing ingredient” may be neural plasticity, since RP may satisfy Hebb’s rule and thereby “differs fundamentally” from feedforward processing.8,109 Accordingly, conscious processing may possibly serve to enable perceptual learning.101,102 Predictive processing & neurorepresentationalism The key idea in PP, and more specifically neurorepresentationalism (NREP) as a theory of consciousness, is that the brain does not accomplish perception by extracting or copying knowledge from its surrounding world but builds up percepts and knowledge by constructing an internal world model. This construction happens by learning to make inferences about the most probable causes of the sensory inputs it currently receives. Thus, “prediction” mainly pertains to inferences about the “here and now,” not necessarily the future. By comparing the effects of internal representations with actual sensory inputs, a prediction error can be computed, which subserves both perceptual learning and inference. When particular circumstances are realized in corticothalamic systems, inferential representations can become sufficiently extensive and enriched to correspond to conscious representations as occurring during perception, imagery, and dreaming.13 As a philosophical position, representationalism localizes perceived objects and features neither in the physical, external world, nor in the brain itself (e.g., if we are conscious of orange, that color is not literally in the brain: neurons do not turn orange). Instead, representationalism attributes these features to the conscious representation we experience as taking shape “outside the brain,” as if “projected” onto the surroundings and our body. NREP starts from this position to identify neural mechanisms underlying the genesis of PC as primary explanandum. NREP takes the computational principles of PP to build sensory-based models of inferential representation learning, positing that motor actions are not needed per se for consciousness.12,13,53 Herein, NREP differs from the active inference account of consciousness, which is also based on PP but does take motor actions (or “agency”) to be at the heart of consciousness.10,110 Below, we will focus on NREP or refer to PP-based accounts in general. Widespread evidence suggests that prediction or representation is a key brain function in realizing a plethora of sensory, motor, and cognitive processes, some carried out consciously and others not.12,110,111 Why are some representations conducive to consciousness but not others? In NREP, conscious phenomenology is characterized by five inalienable hallmarks. Together, these define conscious experience as a multimodal, situational survey of the agent’s world, including its body.12,13,53 This survey or “overview” (in a multimodal sense) subserves, but does not equate to, complex decisions and planning of goal-directed behavior. Multiple modalities contribute to this spatially encompassing survey, not only by way of multisensory integration, but also by modality segregation, such that modalities and submodalities are experienced as qualitatively different. In neural terms, this survey is mediated by multiple sensory-cortical hierarchies (somatosensory, auditory, visual, etc.) that each have a feedforward/feedback architecture and interact with each other, and with motor and planning systems13 (Figure 4). In this very large system, not only the interplay of feedforward and recurrent processing is required for PP in a single modality, but also long-range and lateral interactions between the modalities and, in moving subjects, motor systems.112 Indeed, a major part of the recurrent wave of visual cortical responses to perceived stimuli can be explained as motor-related activity (e.g., Allen et al.113 and Oude Lohuis et al.114,115). This configuration for “superinference” is not found in other brain structures capable of generating predictions (e.g., cerebellum); hence, these are not implied in consciousness. NREP considers several neural implementations of deep PP, varying from strictly hierarchical to heterarchical systems.116,117,118 Under this view, consciousness and attention are different processes: whereas multimodal topology provides for basic conscious experience, attentional systems highlight a limited processing stream within that overall topology. Figure 4 Neurorepresentationalism (NREP) in relation to predictive processing Show full captionFigure viewer A distinguishing aspect of NREP is that conscious representations have a multi-level nature. This particular feature offers an approach to the Hard Problem of consciousness1: the key idea is that “levels” should not only be understood in terms of neuroanatomy or aggregate size (consciousness may be realized in different substrates, even artificial ones). Instead, phenomenology emerges at the top of a hierarchy of conceptually different levels, having single neurons at its bottom and running up higher via assembly-level, unimodal, and multimodal network representations (Figure 4). Dendritic integration theory DIT14,56,120 emerged out of investigations into the role of recurrent processing in the cerebral cortex. It focuses on deep-layer pyramidal neurons of the cortex and the non-linear processes in their dendrites that allow these neurons to serve as the fulcrum, or nexus point, for the convergence of cortico-cortical and thalamo-cortical information flow. It is (currently) based on six core observations, primarily on subcortically projecting (extratelencephalic, ET) layer 5 (L5) pyramidal neurons in rodents: (1) L5 pyramidal neurons associate information arriving at separate dendritic compartments through a non-linear apical dendritic process leading to burst firing.120,121,122 (2) Suppressing these apical dendritic processes (e.g., dendritic calcium spikes) in pyramidal neurons suppresses conscious perception.123 (3) Suppressing the subcortical targets (particularly higher-order thalamus) of L5 pyramidal neurons suppresses conscious perception.124 (4) Loss of consciousness due to anesthesia correlates with a decoupling of the apical from the basal dendrites in these neurons.125 (5) Suppressing higher-order thalamus decouples apical from basal dendritic compartments.125 (6) Neuromodulatory state controls dendritic coupling in L5 pyramidal neurons.125 Taken together, DIT argues that the best explanation for these convergent findings is that the apical axis of subcortically projecting pyramidal neurons serves as the fulcrum for both cortico-cortical and thalamo-cortical information flow, which are both (and simultaneously) fundamental to mammalian consciousness.14,56,57 Thus, it follows that multiple processes associated with the integration of synaptic inputs in the dendrites of these L5 neurons are fundamental for understanding how their output can be so tightly linked to consciousness. The fundamental insight of DIT is that L5 cortical pyramidal neurons both receive and process categorically different information in two distinct regions: the apical and basal compartments that are separated both physically and functionally. The electrical separation between these two compartments, which is already substantial, is accentuated by a high “leak” conductance that allows apical and somatic computations to be essentially isolated from each other. However, it is now well established that these neurons can associate information arriving at both compartments through highly non-linear processes dependent on voltage-sensitive ion channels in the dendrites121,122 and controlled by very specifically targeted inhibition and neuromodulation impinging on subdomains of the dendritic tree.126 DIT hypothesizes firstly that the basal compartment of these neurons receives predominantly feedforward information that relates to the specific cognitive or sensory feature being processed in the column the L5 pyramidal neuron. Second, it proposes that long-range feedback information impinges predominantly on the apical compartment of the same neurons providing information (or context) about processes occurring elsewhere in the brain, including other areas of the cortex. The requirement for neurons with multiple compartments in order to explain information flow in the mammalian brain is a distinctive characteristic of DIT that leads to many corollaries and predictions.14,120,127 For instance, a corollary of DIT is that prior knowledge (e.g., semantic memory), that relates to the associations made with perceptual processes, should be dependent on the influence of synapses connecting to the apical compartment. Evidence that consolidated semantic memories are stored in the connections in layer 1 and their influence on apical dendritic activity was recently obtained in rodents.128,129 Furthermore, the observation that higher-order thalamo-cortical projections sustain the coupling between apical and basal dendrites has been proposed to explain the simultaneous completion of cortico-cortical and thalamo-cortical loops and therefore can serve as the mechanism for (anesthesia-induced) loss of consciousness14 (Figure 5). This, in turn, leads to the hypothesis that directed attention, an aspect of AC, is controlled by the selective maintenance of dendritic coupling in different cortical columns, which in turn can regulate the coupling of pyramidal neurons in posterior cortical regions relating to PC. The regulation of coupling by inhibitory and neuromodulatory input also makes specific predictions about the relationship between brain state and consciousness, including dreaming.130 Figure 5 Dendritic integration theory (DIT) Show full captionFigure viewer Comparing theories across aggregate levels Here, we compare all five theories, discussing their similarities and differences, as well as compatibility and complementarity. All theories propose neural mechanisms that are often overlapping or similar, providing potential points of convergence (for differences, see also Box 3). Neural mechanisms can be described at different levels of organization, from the micro-level of (sub)cellular processes, through the meso-level of within-area circuits, to the macro-level of interactions among brain areas. Although different theories of consciousness have their main focus at different levels, these levels are dealt with as being hierarchically organized, making it possible to describe their mechanistic relations and discern differences and commonalities (Figure 6). Figure 6 A hierarchy of aggregate levels implicated in consciousness by various theories Show full captionFigure viewer Box 3 How to move on with consciousness theories: Remaining differences, criticisms, and approaches The current integrative approach to neural theories of consciousness does not imply that one should be naive about remaining fundamental differences and criticisms. In addition to the differential focus on PC vs. AC, several other key differences deserve highlighting. Whereas GNWT, RPT, PP/NREP, and DIT emphasize the role of information processing, computation, and messaging in generating experiential content, IIT identifies an experience with the cause-effect structure supported by a neural substrate. This requires considering not just the substrate’s actual state but its repertoire of potential states,95 whereas the other theories explain specific conscious experiences as actual states, without taking into account all possible states of the system. By consequence, IIT does not identify consciousness with a particular function (see Tononi and Koch131), while the other theories cast neural underpinnings of consciousness in functional terms, e.g., information distribution (GNWT), learning and feature binding (RPT), predictions (PP, DIT, and NREP), and multimodal situational survey (NREP). However, the fact that these and other theories attribute functional aspects to consciousness does not equate them with computational functionalism (e.g., Pennartz12,13). One general critique that can be exerted on neural theories of consciousness is that, after considering the mechanisms illustrated in Figures 1–7, one may still ask, “… but why would these mechanisms give rise to consciousness”? This question refers back to the “explanatory gap” or “hard problem” that is dealt with by the various theories in different ways (see main text) but should also be taken to heart as serious critique. For instance, the commonly embraced principle of recurrent processing may be necessary for consciousness but is unlikely to be sufficient. Thus, current theories risk being underconstrained and will require further elaboration, based on empirical evidence and theoretical reasoning. A case in point is that a lack of constraints may give rise to differential permissiveness toward consciousness in non-neural substrates (for instance, IIT is relatively permissive; NREP and DIT are more restricted132,133). Moreover, next to the distinction between “structure” and “message,” the question of how messages acquire meaning (intentionality, “aboutness”) remains lurking in the background (e.g., Pennartz,12 Chalmers,15 and Searle134). How to move on through this seemingly intimidating philosophical minefield? One approach is the adversarial testing of theories on points where they make differential, empirically testable predictions. Five such projects have been sponsored by the Templeton World Charity Foundation, including tests of GNWT, IIT, PP/NREP, and RPT. Experiments target, for instance, the anatomic loci representing conscious content, stimulus duration and ignition-like processes,5 and the roles of inactive neurons and motor activity in visuospatial perception (cf. https://www.templetonworldcharity.org/projects-database/0646). However, we argue this adversarial testing will benefit from complementary approaches because it may (1) test predictions that are tangential to the theories under scrutiny, being unable to test their core assumptions, and (2) be over-reliant on falsificationism, whereby a clear-cut reject/accept dichotomy on theories is expected. What appears to be false under one theory may be perfectly acceptable under another one, and improvements of theories may rather follow incremental evidence in favor or against certain accounts, in agreement with abductive inference (“inference to the best explanation of the data”135,136) and Bayesian evidence accumulation.137 We thus advocate a complementary approach whereby new empirical evidence is accrued to combine the best supported elements of various theories into an improved framework with stronger predictive power. Finally, we also expect these combined approaches to gradually push back some of the seemingly deep conceptual (and sometimes metaphysical) divides. For instance, defining more empirically based constraints on theories is expected to restrict the range of complex natural systems to which consciousness may be reasonably attributed. Micro-level: Cellular and subcellular levels A common assumption in much of consciousness science is that consciousness can be best understood at the level of large-scale networks and their computations. However, in the brain global activity is dependent on processes happening at the level of circuits, cells, and even subcellular processes. Here, we take a multiscale approach13,140 trying to unravel some of the processes underlying consciousness, starting from the cellular level (Figures 5 and 6). Recurrent processing and integration—two powerful principles operating at all levels in the nervous system—are both central in most theories of consciousness, including DIT, RPT, GNWT, IIT, and PP/NREP. DIT, in particular, posits that recurrence and integration are inextricably linked through the non-linear processes in the apical dendrites of pyramidal neurons that are a critical nexus for both cortico-cortical and thalamo-cortical loops (Figure 6C). In the awake, conscious state, layer 5 pyramidal cells in cortex are in a “coupled state” so that feedback arriving at the apical dendrites can influence somatic processing at the cellular level, and thus the internal information contained in feedback can be reintegrated into the whole system in a massively parallel fashion.14,125 According to DIT, dendritic integration associated with conscious states ensures that the whole thalamo-cortical system is in a state where activity can be integrated over space and time.56 This could mean that activity can propagate and spread through cortico-cortical recurrent connections, and pyramidal cells can activate higher-order thalamic nuclei, which, in turn, can reinforce the activity in cortex.14,57 According to DIT, the evolving activity gives rise to the rich global dynamics of consciousness measured at the macroscale.130 DIT assumes that such global coordination cannot happen if L5 pyramidal cells are decoupled, e.g., during anesthesia or dreamless sleep130; in this case, activity can propagate locally within the cortex but cannot involve larger thalamo-cortical loops.130 Thus, large-scale brain dynamics (measured by fMRI, EEG, etc.) related to consciousness depend on dendritic integration happening at the subcellular level. In DIT, this local mechanism lays the foundation for processes related to consciousness at other levels of processing.14 For instance, IIT predicts that decoupling the pyramidal cells causes a drastic breakdown in the cause-effect repertoire within the main complex, reducing the Φ value, and hence consciousness.7 Thus, dendritic integration provides a specific cellular mechanism for controlling information integration in the thalamo-cortical system, in line with computational modeling.54 Similarly, if dendritic integration is central for information flow in thalamo-cortical networks, global workspace activity would collapse if the apical and basal compartments of neurons in the frontoparietal network are decoupled.14,125 Thus, sufficient activity propagation and integration in the thalamo-cortical system to cause large-scale ignition may only occur when L5 pyramidal cells are in the coupled state. Similarly, local dendritic integration may enable the build-up of joint neural activity patterns in multiple sensory modalities and across scales that is crucial in NREP. In sum, subcellular dendritic integration is probably an enabling condition for large-scale integration across cortical or thalamo-cortical networks. Thus, the mechanisms proposed by DIT can be incorporated in any theory of consciousness that relies on such large-scale integration. Dendritic integration also provides neurobiological nuances to theories of consciousness that have a dual-stream architecture (such as RPT and classical PP; cf. Rao and Ballard141). In particular, the knowledge about dendritic integration mechanisms specifies how the two streams interact at the level of single pyramidal neurons. According to RPT, consciousness arises from recurrent interactions between different levels of processing in the brain. This idea is compatible with DIT as decoupling of pyramidal neurons leads to a breakdown of RP in the brain.14,125 According to PP-related accounts of consciousness (e.g., NREP), predictions sent from higher to lower levels of processing try to “explain away” activity propagating from lower to higher levels by suppressing the sensory signals that match the predictions. What cannot be predicted is passed on to higher levels as prediction errors. Although, according to DIT, the biophysical processes intrinsic to the L5 pyramidal cells prescribe them to amplify the match between the two streams,14,120 there are multiple ways in which the biophysical properties of pyramidal neurons, or adjoining cortical interneurons, may be co-opted into a PP framework (Figure 6C135,142,143,144,145). All of these variants are compatible with principles of dendritic integration as a cellular basis used to generate conscious experience. However, according to DIT, recurrence is only one requisite for consciousness. Another key component is the dendritic integration mechanism that enables the incorporation of this feedback information into the ongoing activation patterns of the thalamo-cortical system. Meso-level: Local circuits and recurrent processing How may local processing within and between corticothalamic microcircuits contribute to consciousness? Besides the large L5 neurons highlighted in DIT, multiple neuron types connected by millions of synapses form immensely complex microcircuits within and across cortical columns,146 the intricacy of which is easily underestimated. This dense complexity, multiplied by similarly rich inter-areal and cortical-subcortical interactions, may transcend human intuition and may (unpredicted and unexpectedly) contribute or lead to the seemingly non-physical phenomena of subjective experience (PC). The local, nested intricacies within and between cortical neurons and microcircuits are relevant for all the five theories discussed here and may, e.g., contribute to high Φ cause-effect structures underlying P-consciousness according to IIT.95 In general, feedback (RP), both within microcircuits147 and more globally, is an important feature in most consciousness theories (RPT, GNWT, IIT, PP/NREP; see macro-level: multi-area systems and overarching concepts: richness and complexity) and also contributes to processes linked to consciousness, such as WM (Figure 3) and sensory adaptation.20 However, what could local processing contribute specifically to consciousness? Building on the DIT-based predictions in micro-level: cellular and subcellular levels, PP/NREP has proposed that the computation of low-level predictions and prediction errors strongly depends on the local, intracolumnar circuitry that regulates information trafficking in L5, but also L2/3 pyramidal cells. Accumulating evidence suggests a role for superficial pyramidal cells in representation (prediction) and error coding, while L5 cells may function especially in representation coding.23,118,135 In particular, somatostatin-positive interneurons have been implicated in error computations.148,149 Although RPT focuses primarily on RP within and between visual areas,150,151,152 Lamme also discussed whether even RP between interconnected neuron pairs may perhaps generate some consciousness. However, the ubiquitousness of feedback led Lamme to propose that some other process, perhaps neural plasticity, may also be needed for PC.102 GNWT implies that conscious access (AC) depends on RP, as its “ignition” and reverberant activity realizes access to conscious content.28 Early versions of GNWT proposed that L2/3 pyramidal cells with long-range connections (not L5 neurons as in DIT) are most crucial for AC, but more recent versions also consider the loops with subcortical structures. The Dehaene et al.’s model of AC,153 which included thalamo-cortical columns with laminar feedforward and feedback projections, simulated ignition with a P300 wave and attentional blink. The network model154 proposed how a short-lasting (<1 s) perceptual buffer can be accessed and broadcasted to WM and consciousness (AC), by distinct operational modes of the same circuit. According to IIT, feedback (RP) is needed for consciousness. Based on how Φ is calculated to “quantify” PC, artificial systems with purely feedforward architecture have Φ = 0, and hence lack consciousness (PC), even when their overt behaviors (input-output functions) are indistinguishable from that of conscious (Φ > 0) systems with feedback94,95 (but see Doerig et al.155). Accordingly, consciousness should fade even in the presence of the appropriate architecture if feedback connections are functionally disabled. On the other hand, since IIT holds that the causal structure of brain circuits, rather than their activity, determines Φ and PC, circuits that are causally integrated within the main Φ complex can contribute to PC, even when its neurons are silent. This counterintuitive prediction, not shared by other theories, is being tested in an adversarial collaboration between IIT and PP (see: www.templetonworldcharity.org/projects -database/0646; Box 3). Haun and Tononi propose that “grid-like” neuronal circuits in sensory cortices, resembling lattices, form a substrate for spatial experience.97 Grid-like connectivity is found in both visual and other sensory areas, but PP accounts generally interpret the topographic structure of sensory cortices differently than Haun and Tononi, viz. as a layout for making inferences from one part of sensory (e.g., visual) space to another part (cf. Knierim and Van Essen156). In addition, NREP proposes that subjective spatial experience arises from the multimodal integration between maps set in different spatial frameworks (retinotopic, craniotopic, and allocentric12). In addition to representing information content, local circuits undergo changes in neuromodulatory state with different “levels” of arousal (reflecting one of the main “classical” “dimensions of consciousness157”; but see Hill and Tononi,158 Laureys et al.,159 and Bayne et al.160). However, according to a more recent view of neuromodulatory states, these can be regarded as “background conditions” or enabling factors with only an indirect relation to consciousness per se (not specifying conscious content).16 The effects of changes in local “arousal” state and network complexity caused by anesthesia or neuromodulation have been modeled in thalamo-cortical network models158 and seem roughly compatible with all five theories discussed here, although dendritic mechanisms (DIT) were not included so far. Such modulatory effects have been assessed even in isolated cortical slices (“meso-scale36”) by modified versions of perturbational complexity index (PCI) and related complexity indices similar to those used in humans (see Box 1; overarching concepts: richness and complexity). Although these measurements do not allow any inference about consciousness in brain slices and rodents, they can inform about cellular and network mechanisms responsible for brain state changes with loss and recovery of complexity.161 Macro-level: Multi-area systems When ascending along the scale of neuroanatomy, we arrive at the question of how brain mechanisms underlying conscious experience operate at the level of brain areas and systems composed of various areas. The five theories reviewed here can be subdivided into three partitions on this point. First, IIT, PP-NREP, and RPT have in common that, for conscious vision, they all emphasize the foremost role of the posterior cortex (more specifically, those occipital, parietal, and temporal areas implied in vision). IIT holds that a maximal Φ value will be reached primarily in posterior cortex (“hot zone”)78,95 when dense reentrant interactions occur between higher and lower visual areas.7 PP-based accounts, including NREP, propose that conscious vision begins with the interaction between the feedforward signaling stream (mainly conveying error information) and the recurrent stream (conveying top-down predictions) within the visual cortical hierarchy,23,141,135 although alternative schemes are possible (e.g., Heeger162). Both IIT and NREP propose that, in a changeable visual environment, different, switchable subsets of brain areas will engage in feedforward and feedback interactions to ensure integration of visual features, depending on the visual submodalities being part of the subject’s experience.12,88 RP within the visual cortical system is also key to RPT in reaching PC, while global recurrent processing is needed to accommodate attentional effects and reach AC. GNWT attributes a prominent role to the prefrontal and parietal cortices in conscious vision, which is in line with its focus on AC rather than PC, and its translation of this focus into executive functions such as behavioral responding, attention, WM, and valuation—all functions associated with the prefrontal-parietal network.6 Third, as we have seen, DIT focuses more strongly on the cellular basis of consciousness. In that sense, DIT is “neutral” with respect to the “anterior versus posterior” debate on neural substrates of conscious vision. Notably, all five theories agree that the neocortex constitutes the key structure for generating conscious content (as opposed to other theories highlighting centrencephalic structures163,164,165) and that thalamic nuclei are vital in sustaining consciousness. Moreover, all five theories posit that, next to local or short-range interactions, long-range communication in corticothalamic systems is required for consciousness.6,7,13,105 Importantly, however, both RPT and NREP argue that information processing within the visual cortical hierarchy is not sufficient to achieve conscious vision (either understood as PC or AC). According to RPT, widespread recurrent processing requires attention, and for this, interactions between the visual cortex and prefrontal areas are deemed necessary.102 In this sense, RPT is more compatible with GNWT than their different definitions would suggest: they both attribute a key role to the prefrontal-parietal network regulating attentional signals.102,105 PP accounts interpret attention in relation to large prediction errors and low decision confidence, prompting the subject to seek more information about perceptually uncertain regions of sensory space (e.g., Kanai et al.166). Among the hallmarks of conscious experience delineated by NREP, we find the combination of dynamics and stability, where “stability” refers to the capacity to perceive the outer world and its objects as a stable whole, despite the occurrence of eye, head, and body movements.12,13 To achieve visual stability in physically active subjects, interactions between the visual and motor cortices are required: the visual system receives motor signals (along with proprioceptive and vestibular reafference) or predictions about which sensory changes are happening or about to happen.112,166 Moreover, NREP postulates that lateral, multisensory interactions between sensory-cortical systems are necessary to ensure the qualitative, multimodal richness that characterizes consciousness.167 By consequence, vision is not mediated exclusively by the visual corticothalamic system as classically defined but by the “extended visual cortex,” which includes many satellite areas. IIT and RPT are inter alia, not in contradiction with these additional interactions, but do not make a particular point about this. Thus, at first sight, three theories appear to agree at least globally on the macro-anatomy of conscious vision, while the other two theories either express a different view—in relation to differences in background definition (GNWT)—or stay neutral (DIT). This global compatibility of IIT, RPT, and NREP does not mean they have the same scope or are similar. For instance, they use different rationales as to why recurrent feedback in posterior cortex would be required for consciousness (IIT: to maximize Phi; RPT: to enable plasticity, learning, and feature binding; NREP: to close the loop for representation learning and inference). Another key difference lies in the presumed role of inactive neurons (in IIT, these can contribute to conscious experience unless they are disconnected from the network or inactivated55; other theories do not adhere to this). Nonetheless, we may conclude that, also at the level of multi-area systems, the theories are not as incompatible as has sometimes been assumed. Overarching concepts: Richness and complexity A growing body of experimental evidence suggests that the patterns of neuronal interactions that are relevant for consciousness must not only be tightly integrated through feedforward and feedback connections but must also be differentiated and information rich.32,168,169 This notion of brain complexity—defined as the coexistence of integration and differentiation—was at the core of IIT’s original notion of Φ89,90 and remains central to the empirical predictions of IIT in its mature form.95 Brain complexity has also been embraced by other theories more recently. For instance, a key tenet of the GNWT is the ignition of widespread activations. Hence, IIT and GNWT share in practice the basic requirement of neuronal integration, albeit motivated by divergent theoretical underpinnings. Notably, the original formulation of GNWT made no explicit reference to information-related measures, differentiation or complexity and mainly focused on neuronal processes such as fronto-parietal synchrony and the P3b. Subsequently, however, empirical studies by GNWT proponents shifted the focus to the richness of dynamical repertoire,30,170 and, more recently, the notion of complexity has been explicitly incorporated as a key element within the GNWT, based on the rationale that distributed modules broadcast different types of sensory information.171 As defined by IIT, brain complexity bears clear mechanistic relationships with the microscopic and mesoscopic processes postulated by DIT and RPT. It requires effective reentrant interactions to ensure that differentiated groups of neurons cohere tightly and engage in a rich repertoire of cause-effect interactions. Both dendritic decoupling and the breakdown of recurrent interactions are expected to result in dramatic decreases in the brain’s capacity for integrated information. Along these lines, the collapse of complexity observed experimentally during sleep and in unresponsive wakefulness syndrome (“vegetative” state) patients has been associated with a breakdown of feedback interactions.37,172,173,174 Similarly, complexity is inherent to the multimodal richness and integration, postulated by NREP, although there are also important differences. In IIT, brain complexity explains general phenomenal properties (unity and richness) that are “immediate” and common to each and every experience, while both NREP and PP in general take conscious experience to be representational—a best guess “simulacrum” of external reality.12,13,175,176 In the neurorepresentational view, complexity translates into a multifaceted phenomenon comprising multimodal, qualitative richness of conscious content, spatial extendedness and resolution, intensity and attention, and finally temporal depth—all of which are understood to be conducive to subsequent action planning.13,53 In this respect, there are many ways of defining and operationalizing brain complexity, and more efforts are warranted to specify underlying assumptions and extend notions of complexity into an exhaustive taxonomy.32 Nonetheless, the recent convergence on the general notion of complexity88 has interesting implications—especially for empirical applications in unresponsive patients170,177,178 (Box 1). Widespread theoretical agreement may increase such confidence and justify inference in more challenging cases. Furthermore, complexity and information richness can be studied at multiple levels (Figure 6), from human extracranial and intracranial recordings to rodents38,169,179,180,181 and cortical slices.182 Outside of the human brain, in which they are calibrated, current complexity measures do not allow any inference on the level of consciousness. Yet, they offer the prospect of drawing a mechanistic link between neural processes occurring at different scales, from microscale dendritic integration to mesoscale recurrent processes and macroscale inter-areal integration. Finally, this convergence around complexity may represent a common background to contrast different frameworks at a finer granularity. Where and how in the brain is complexity instantiated? For IIT the cause-effect structure of grid-like cortices with recurrent pyramid-like, convergent-divergent connections (such as those found in posterior areas) (Figure 7B) is thought to correspond to a maximum of complexity and thus be sufficient for conscious experience.96,97 PP-NREP shares with IIT the quest for the neural basis of the percept of spatial extendedness but seeks this in the integration between retinotopic, craniotopic, and allocentric frames of reference used by the occipital, temporal, and parietal cortices12; these different directions may prove to be compatible. Another key question is: why do some theories of consciousness focus on the complexity of ongoing neuronal activity, while others emphasize the complexity of the underlying causal structure?183 It will be important to clarify this aspect, as it entails not only different approaches to measurement (observational vs. perturbational) but also a substantially different understanding of the kind of information that matters for consciousness (an extrinsic message vs. an intrinsic cause-effect structure) (Box 3).7,184,185 Finally, how do current theories consider the possibility that analyzing the structure and complexity of causal interactions within the brain may also be useful in the search for content-specific neural correlates of consciousness? Addressing this question is challenging, but attempts have been made in this direction.97,167,176 Figure 7 Comparing theories of consciousness Show full captionFigure viewer Overall assessment: Toward an integrated theory of consciousness? Is it realistic to reconcile theories, or even aspire to a unified theory of consciousness? At first sight, as we have seen, at least some of the theories seem to be so different (Box 3) that bridging them appears insurmountable. This has even led to accusations of illegitimacy between competing theories.187 However, we take the standpoint that the existence of multiple theories is a sign of healthiness in this nascent field and that paying close attention to the explananda can disentangle the core claims such that multiple theories can simultaneously contribute to our understanding. Fundamental differences in the focus and explananda of theories may sometimes make them more compatible. For example, GNWT focuses on AC, whereas RPT, IIT, and PP/NREP aim to explain primarily PC. An integrated theory of consciousness will have to explain PC in neural terms but also how conscious content is distributed and used for subsequent action planning, WM, valuation, attentional regulation, etc., which is the domain of AC. Modeling efforts in this direction are indeed being made, for instance, to integrate PP accounts with global neuronal workspace operations.188,189 We also saw how RPT can be integrated, in principle, with GNWT in a joint effort to include AC. As IIT already acknowledges the role of prefrontal areas for access, there is no contradiction with GNWT regarding its “anterior” focus for AC. Still, some will argue that GNWT and IIT cannot both be right because they disagree on the nature of what consciousness is (see also Box 3). A further point of potential conceptual convergence lies in high-level, complexity-related concepts such as information integration (IIT), global information availability (GNWT), multimodal richness in multi-level representations (NREP), recurrent processing at different anatomical scales (RPT), and integration in corticothalamic loops (DIT). Although, at face value, these ideas differ remarkably, they are all fundamentally based on different aspects of interconnectedness, recognizing that the richness of neuronal interaction, relying on short- and long-range connectivity in corticothalamic systems, is an important factor for significant degrees of consciousness (Figure 7). These points of convergence and related considerations give reason to fuel a certain hope that eventually an integrated theory of consciousness—or at least a related set of interdigitating concepts and working hypotheses—may come within reach within the next years or decades. However, it must also be recognized that there are still vast areas of theoretical and empirical investigation that remain to be unlocked (Box 3). The realization that we are only scratching the surface of this great question is humbling, but consistent progress is now apparently being made on several fronts. The state of consciousness research at the moment may be compared with a multifaceted object examined from different angles, each stressing different aspects based on individual vantage points. A parallel can be drawn with quantum mechanics, where there is also agreement on observations and phenomena, but still widespread disagreement on how to interpret the relationship between probabilistic, micro-level descriptions and our deterministic, macro-level everyday lives. Both cases may possibly reflect fundamental limitations of human imagination and intuition.12,190 This analogy underscores the urgent need for consciousness researchers to develop well-defined concepts and a common terminology (e.g., not use merely the general term “consciousness” for different concepts like “AC” and “PC”) and to devise ways to translate disparate formulations into one another. Just as quantum theory benefited from a robust mathematical framework, so too should this field. Finally, we feel there is broad and uncontroversial support for more experimental testing and development of empirical measures of consciousness. The relevance of this is 2-fold. First, novel, theory-based empirical measures such as PCI and other complexity-related measures can be usefully applied in clinical settings, such as in helping to classify unresponsive patients with consciousness disorders into different subgroups with more nuanced distinctions and different prospects for recovery and treatment (e.g., Demertzi et al.,170 Casarotto et al.,177 and Farisco et al.191). Second, empirical measures may help to distinguish or unify different theories. The high-complexity spatiotemporal EEG patterns evoked by TMS in conscious conditions are invariably associated with the presence of recurrent waves of activation,174,192 whereas low-complexity patterns are found during cortical up and down states characteristic of slow-wave sleep and anesthesia (e.g., Steriade161 and Olcese et al.193). Hence, PCI is compatible not only with IIT postulates but in principle also with the neural processes that are relevant for GNWT, RPT, NREP, and DIT.13,14,102,171 The field is already benefiting from a broader palette of empirical tools and measures. DIT, for instance, introduced EEG and LFP markers that align with dendritic Ca2+ spikes during distinct consciousness phases.194 PP/NREP proposes representational measures of consciousness (relying on techniques to decode conscious content that subjects are experiencing; cf. Goltstein et al.,195; see also Box 2), and GNWT stands to benefit from report vs. no-report paradigms combined with single-unit recordings, fMRI and EEG-MEG (e.g., Sergent et al.42). Complementary computational approaches can be utilized to foster further comparison and integration,116,188,196,197,188,196,197 and research is being extended from human to animal consciousness to benefit from in vivo two-photon imaging, large-scale ensemble recordings, complexity measures, and optogenetics to unravel causal substrates of consciousness (e.g., van Vugt et al.,64 Oude Lohuis et al.,114,115 Takahashi et al.,124 Suzuki and Larkum,125 Arena et al.,38,179 O’Connor et al.,198 and Storm et al.199). Adversarial research is emerging as an important new approach in consciousness research,5 but it remains to be seen whether this will eventually lead to a Darwinian selection of a single, winning theory (Box 3). Here, we have advocated for a more unifying, integration-oriented approach, which seeks to empirically test and combine valuable elements from various theories. In conclusion, we have argued that neural theories of consciousness (1) show differences that are often not as large or insurmountable as initially thought, (2) offer opportunities for combining and integrating various elements figuring at different levels of organization, and that (3) an integrated framework for understanding brain-consciousness relationships, guided by both empirical and theoretical advances, has strong potential to move the field forward in the next decade. translate above into Chinese. Do not just summarize
- Assistant
摘要 意识体验与物质大脑过程之间的关系如何?近年来意识研究的激增催生了多种理论来回答这个古老的问题,其中一些理论目前正处于激烈的辩论之中。尽管大多数研究人员迄今为止主要集中于相对孤立地发展和验证他们所偏好的理论,但本文由一组代表不同理论的科学家撰写,采取了另一种方法。我们注意到,各种理论往往试图解释意识的不同方面或机制层次,因此我们认为这些理论并不一定相互矛盾。相反,它们中的几个可能在基本神经机制上趋于一致,并且在某种程度上是兼容和互补的,因此多个理论可以同时为我们对意识的理解做出贡献。在这里,我们考虑到尚未得到充分重视的统一、整合导向的方法,试图结合各种理论中的有价值元素。
引言 比较理论 心脑问题,即我们的意识体验如何与物质大脑过程相关,几个世纪以来一直是哲学家们争论的话题,至今仍然是科学中最深奥的未解之谜之一。过去几十年,意识研究的理论和实证研究激增,提出了多种神经科学理论,并引发了激烈的辩论(例如,Seth和Bayne)。为了测试竞争理论之间的差异,Templeton世界慈善基金会赞助的“对抗性合作”已被开展(例如,Cogitate Consortium等)。
历史上,大多数研究人员强调孤立地发展和验证他们所偏好的理论框架。然而,本文采取了不同的方法。它源于人类大脑计划(HBP;2013-2023)中的研究人员,他们在与意识相关的主题上工作,并代表不同的理论。基于最近的实证进展,我们表明,许多意识理论的不同方面并不一定相互矛盾,正如有时所声称的那样;相反,这些理论往往试图解释意识的不同方面,并趋向于在基本神经机制和过程上达成一致。在这里,我们主张,几种理论至少在某种程度上是兼容和互补的,我们考虑到尚未得到充分重视的趋向整合的方法。
由于无法涵盖所有提出的基于神经科学的意识理论,我们在此重点关注五种我们认为代表显著和互补视角的理论:(1)全球神经工作空间理论(GNWT),(2)整合信息理论(IIT),(3)递归处理理论(RPT),(4)预测处理(PP)和神经表征主义(NREP),以及(5)树突整合理论(DIT)。这些理论涵盖了Seth和Bayne所概述的四个广泛类别中的三个,因此允许广泛讨论不同理论如何可能被调和和整合。在这里,我们使用“理论”一词是广义的,因为没有公认的定义(表1),而其他术语(“假设”、“模型”和“框架”)也缺乏严格的定义。我们关注这些理论和思想的内容,而不是它们的状态或成熟度。选择理论的另一个标准是每个理论都应代表一个累积的神经科学研究体,汇集在一个可定义的思想框架下。
表1 术语和概念
(表格内容省略)
通过检查这五种理论的比较及其可能的部分调和,我们旨在启动一个整合过程,未来也可以纳入其他理论的有价值元素。然而,某些当前的理论当然可能被证明是完全或大部分错误或无效的,并将在未来的选择中被剔除,正如科学历史上反复发生的那样。为了克服长期以来似乎无法解决的哲学和理论辩论,我们提出了一种不同的方法,专注于基于实证的整合、复杂性、表征和不同层次的递归处理的概念,这些似乎是这些理论所共有的特征。
根据Ned Block的观点,我们区分了现象意识(PC;即立即的、主观的体验,例如看到蓝天)与可访问意识(AC)和可报告性机制。我们讨论了这些区分与提出的意识实证测量之间的关系(框1和框2),以及它们与注意、期望、计划行为和不同形式的记忆之间的关系。我们还区分了从低层次的机制特征到高层次的系统性属性的解释层次。
框1 评估意识状态和层次的方法 识别基于大脑的测量以可靠地标识人类意识的存在与否是一个重要的文献焦点,对于临床医生来说也是一个基本关注点,且可以说是理解意识与其物理基础之间关系的重要第一步。在实践中,当意识的行为评估不可靠时,基于大脑的测量是必要的,例如在因全身麻醉而瘫痪或因严重脑损伤而无反应的受试者中。
第一类标记涉及绕过运动行为,直接测量对感官刺激的脑反应,例如语言命令或复杂的声音序列。一个显著的例子是P3b,它在听觉序列中对全球期望的违反产生反应。根据GNWT,P3b标志着感官信息向前额皮层处理器的广播(在点燃之后),并已被用于检测脑损伤患者的意识存在。尽管P3b提供了一个特定的标记(即仅在意识到刺激的受试者中存在),但在意识的脑损伤患者和健康的意识受试者中通常缺失。因此,P3b的低灵敏度是一个实证提醒,表明AC(标识与环境的连接和/或高层认知功能的参与)与PC的神经相关物之间可能存在解离。
第二类标记直接关注脑活动,而不一定依赖于感官处理和认知功能。最近的实证收敛集中在整合和区分的测量上,这些测量有望更准确地检测意识。值得注意的是,这种收敛涉及许多独立研究者支持不同的理论框架,并在各种条件下使用不同的技术,例如睡眠、做梦、麻醉、癫痫、幻觉状态和昏迷。一般来说,可靠地独立检测意识的前景与对外部环境的连接、行为表现、认知和记忆能力无关,具有临床意义,并促进对PC神经机制的探索。复杂性测量依赖于对自发大脑动态的特征化(通过功能连接网络和图论指标的分析)或对直接皮层扰动所引发的交互模式的量化。在各种生理和病理条件及实验模型中,意识的丧失和恢复期间PCI的减少和增加始终与激活的递归波的消失和复苏相关。
框2 评估意识内容的方法 在细分类别层面上实证评估意识内容是具有挑战性的,因为需要识别代表特定特征的记录位置。Koch等人提出了一种有用的方法,寻找意识的内容特定神经相关物(NCC):(1)映射步骤,在此步骤中,使用适当的特征特定刺激确定记录位置(例如,神经记录研究中的神经元或fMRI研究中的体素)的反应特性;(2)应用一种实验范式,利用多稳定图形或双眼竞争,其中物理刺激特性保持不变,而意识内容在刺激的不同解释之间交替。如果经历的(报告的)意识内容在步骤1中映射的特征动态活动变化中反映出来,则相应的记录位置符合潜在的NCC;(3)实验上尝试排除已知的混淆因素,例如,与意识感知共变的神经活动可能反映意识的先决条件或后果。因此,必须记住,这种方法是相关的,而不是因果的,注意和记忆效应可能会混淆结果的解释。
这种方法已在猕猴和人类中成功应用,使用电生理记录。人类的颅内(立体脑电图)记录导致识别出音调活动模式,即对感官刺激的延长低幅伽马波反应,可能对应于反映感知意识的先前报告的晚期EEG成分。此外,视觉-触觉刺激特别增强这些音调成分,平行于视觉-触觉刺激显著改善触觉消失患者的检测的概念。
该方法可以扩展到具有足够高空间分辨率的人类fMRI测量,以区分皮层柱级别的特征特定反应。在Schneider等人的一项研究中,选择性响应于水平与垂直运动刺激的人类中间颞叶复合体(hMT+)中的柱状簇在映射阶段被定位。在主要实验中,呈现了一种模糊的明显运动刺激(运动四重奏),可以被感知为水平或垂直运动。通过跟踪在预映射的水平/垂直柱状簇中的活动,可以预测参与者在交替的时间段内感知的两种可能解释(水平或垂直运动)。运动四重奏研究的结果支持hMT+是(部分)方向运动的内容特定NCC的观点,这与Salzman和Newsome在非人类灵长类动物中阐明的中间颞叶(MT)在运动视觉中的因果作用一致。更一般地说,映射皮层(柱状)功能簇并使用亚毫米超高场fMRI在模糊感官刺激期间测量其反应,满足Chalmers和Koch等人概述的寻找内容特定NCC的标准。
概念与解释 在比较理论时,首先澄清关键概念、定义所使用的术语,并识别理论旨在解释的现象(解释对象)是至关重要的。通过我们的比较方法,立即突显出不同理论往往关注“意识”的不同方面或亚类型,这可能导致概念混淆和误沟通。对立的理论可能会促进不同的实验方法,这些方法并不直接可比,或者对数据的不同解释。例如,GNWT专注于AC,主要试图解释意识内容如何被主体访问和使用,以及意识与无意识的“信息处理”(而不是现象体验),通常研究对特定感官输入的反应,例如看到一张面孔。相比之下,整合信息理论(IIT)、递归处理理论(RPT)和预测处理(PP)/神经表征主义(NREP)主要关注“PC”,即体验某种事物的“感觉”,无论其是否可以报告。因此,与GNWT相反,其他理论(IIT、RPT和PP/NREP;对于树突整合理论(DIT),见下文)主要试图解决“硬问题”,即解释如何可能在像大脑这样的物理系统中产生PC和“质感”,以及为什么这些主观现象在无梦睡眠或麻醉期间消失(但见Sevenius Nilsen等人)。
这种差异部分源于不同的哲学观点:与IIT、PP/NREP或RPT相反,GNWT的支持者通常不认为PC或质感是一个真实的现象,与AC不同,而是认为只要他们的理论解释了可观察的事实(例如,报告:“我现在看到脸!”),就没有什么需要解释的。对于PP/NREP、IIT和RPT,PC是主要的解释对象,而对意识内容的访问则被视为附属事项,依赖于将意识内容应用于运动或认知操作的执行过程。DIT则提出了AC和PC的机制。
GNWT与IIT的原始版本在另一个方面也有所不同:GNWT主要关注意识内容的选择和使用过程(AC),而IIT则更关注PC的基本性质和特征(在信息整合和交互方面),以及PC的状态和水平。然而,现在这两种理论越来越多地试图解释过程、内容和状态/层次,这一趋势与RPT、PP/NREP和DIT共享。越来越多的共识认为,任何完整的理论都应同时解释意识的内容和状态/层次。最后,DIT是一种最近提出的理论,主要在细胞和电路层面上运作,并解释为什么哺乳动物的意识依赖于皮层-皮层和丘脑-皮层环路的同时完成。它假设AC和PC都涉及由于树突处理而同时完成这些环路。
在本综述中,我们将首先简要介绍这五种理论,然后讨论它们之间的比较,使用大脑的层次组织作为支架,接下来考虑不同理论是否以及如何可能被调和和整合。
意识理论 全球神经工作空间理论 全球工作空间理论(GWT)描述了产生意识体验的处理步骤,作为局部处理器与全球工作空间之间的动态交互,全球工作空间整合信息并将其广播到其他处理器以进行进一步处理。因此,全球工作空间充当了一个骨干,通过它不同感官和认知过程的信息变得可供多个大脑区域使用。正是广播的信息构成了意识的内容。GWT的一个神经生物学启发版本后来被提出为GNWT。它采用了全球工作空间的相同功能概念,但将其与特定的大脑网络和神经特征联系起来。在这里,局部处理器包括具有感知、行动、记忆等功能特化的模块化皮层区域。它们的活动由感官刺激触发,并且可以相互独立地进行计算。例如,这些特征在早期感官皮层中被发现,其中不同特征在专业的皮层区域中被处理。全球工作空间被认为涉及长程连接(皮层-皮层和丘脑-皮层),连接广泛的大脑区域,特别是在前额皮层、后顶叶皮层和相关区域。这些分布区域可以共同选择一个心理对象,以便增强和绑定该对象的所有相关特征的表征。连贯的对象表征的形成对应于感知心理学中的基于对象的注意。在这种方案中,丘脑和其他亚皮层结构(如基底神经节)也可能用于放大和分发信息,并整合大脑各大区域的神经活动。皮层区域被认为选择和整合信息,以使大脑能够专注并执行复杂的认知任务。最后,当局部神经元集合达到进入全球工作空间并参与后续广播的活动阈值时,称为“点燃”。
点燃是GNWT中的一个核心概念,已与递归处理联系起来,例如在前顶叶皮层中。近年来,它得到了支持,例如来自人类神经生理学和神经成像研究,这些研究显示刺激出现后200到300毫秒内大脑活动的明显分歧,这取决于意识感知。意识感知与晚期事件(例如N2/P3b/P300成分)的关联,已在不同感官模式和范式中观察到。早期感知处理在前200毫秒内可能在意识(即报告)和无意识(即未报告)试验之间相似或甚至相同,但意识感知与晚期事件相关,例如人类头皮事件相关电位的N2/P3b/P300成分。意识访问的时机在某些条件下也可能延迟,例如在注意力不足或双重任务情况下,此时全球工作空间已经被另一个意识对象占用。人类前额叶的概念细胞和非人类灵长类动物的前额皮层神经元在相应刺激被意识到时表现出更强和更持久的放电率。在猴子中,信号传播到前额皮层进一步导致点燃和行为报告。这些实验结果中的一些存在争议,通常基于感知报告可能混淆后期前额活动的观点。然而,最近在猴子中无报告范式的结果可能在一定程度上缓解了这些批评。
大脑区域内或跨区域的递归电路支持持续活动和类似点燃的过程。全球工作空间所需的长程皮层投射起源于皮层的II/III和V/VI层,在这些层中反馈效应也最为明显。最近的证据表明,持久神经活动的回声环路涉及皮层-皮层相互作用和丘脑及小脑核团等亚皮层区域,这与支持全球工作空间的广泛长程连接网络一致。
整合信息理论 整合信息理论(IIT)强调体验的“现象结构”,使用内省来表征现象属性,然后根据相应的物理属性推导这些属性。这个物理解释可以根据测量行为、功能和神经相关物的标准方法进行实证验证或反驳。意识优先的方法的种子可以追溯到1998年发表的早期论文《意识与复杂性》。在这篇论文中,作者从内省和每个意识体验既是整合的(单一的)又是高度信息化的前提出发,假设每个意识状态以特定方式与大量替代状态不同。然后,他们假设神经整合和区分的共存是意识系统的一个关键物理特性,并将这一主张操作化为可测量的神经特性: (1)整合:一组神经元只有在其成为一个功能簇的一部分时,才能对意识体验做出贡献,该功能簇通过数百毫秒的再入交互实现高整合。 (2)区分:维持意识体验至关重要的是,该功能簇必须能够具备大量不同的神经状态,伴随高复杂度值。
这一原则将意识与神经交互的复杂性联系在一起,是IIT早期表述的核心,激发了“弱IIT”的版本,并导致可测试的解释和预测: (1)一些具有适当解剖要求(高密度的侧向连接和大量的再入电路)的丘脑-皮层系统支持意识体验,而其他具有模块化或前馈架构的结构(如小脑)则不支持。 (2)同步或广泛的神经激活不一定与意识状态相关,除非这些状态具有高水平的区分或复杂性。 (3)鉴于解剖要求得到保留,影响再入、整合和区分的功能变化将导致复杂性和意识的丧失,例如在睡眠、全身麻醉和广泛性癫痫发作的情况下。 (4)即使在没有感官输入、运动输出和执行功能的情况下,内在的大脑交互也可以支持高复杂性和意识体验。 (5)意识的存在与否可以通过使用直接皮层扰动和记录来可靠地评估,以测量大脑内因果交互的复杂性。
这些早期预测在过去20年中已在不同意识状态(清醒、睡眠、做梦、迷幻状态、癫痫和意识障碍)中通过各种脑成像方式进行了广泛评估,并显示出临床应用的前景(有关综述,见Sarasso等人;另见框1)。IIT的完整表述考虑了意识的五个现象属性(内在性、信息、整合、排除和组成),然后将它们表述为物理属性。通过其数学形式,IIT旨在解释意识的存在与否,以及其质量和内容。具体而言,IIT主张每种体验——特定的“现象结构”——与物理基质所指定的相应“因果-效果结构”之间存在解释性同一性。
IIT认为,因果-效果结构捕捉了系统在特定状态下的全部因果能力——例如,一组神经元的网络,有些处于活动状态,有些处于非活动状态。测试这一框架需要结合系统扰动、观察和分区,以精确评估子集元素之间的相互影响(图2)。
递归处理理论 递归处理理论(RPT)的核心思想是递归(反馈、再入)处理对意识体验至关重要,甚至可能是充分的。Victor Lamme基于对灵长类动物视觉的实验发展了RPT,形成了一个视觉意识的通用理论。RPT的一个关键要素是将注意力和意识视为独立的(正交的)过程。这使得RPT与许多其他理论(包括GNWT、PP和注意力模式理论(AST))区分开来,但与NREP没有区别。
RPT概述了视觉处理的四个阶段:两个是无意识的(阶段1和阶段2),另外两个(阶段3和阶段4)被视为意识的,但不同(现象性与AC;PC与AC)。RPT还将PC和AC与注意力和不同形式的记忆联系起来,因此开创了将RPT与其他理论调和的尝试。
阶段1 局部(“表层”)前馈处理:快速的前馈活动(FFS,约100-200毫秒)将信息从初级视觉皮层传递到更高的视觉皮层,在那里提取图像特征。然而,如果没有得到注意的增强,处理将保持局部,无法进入阶段2,仅产生“潜意识”处理,而没有PC或AC。然而,“非前馈”(递归/侧向/水平)处理可以通过FFS在所有皮层水平上引发,从而触发阶段3和阶段4。
阶段2 更全球化的(“深层”)前馈处理发生在刺激被注意时。如果是这样,增强的FFS会传播到运动和前额区域,导致快速(约200毫秒)识别主要内容。这可以触发即时的无意识(“自动”)行为,例如击打网球,同时主体对刺激和行为都没有意识。
阶段3 局部(“表层”)递归处理是一个中间处理阶段,其中未被注意、未被遮蔽的刺激,主要在腹侧视觉通路中,导致现象意识体验(PC)。然而,如果没有注意,递归处理是有限的,无法与前额和运动区域进行交互。有限的递归意味着感官刺激不会进入工作记忆和进一步的认知过程,这些过程是报告所需的。因此,存在PC而没有AC。
阶段4 在充分的时间内,注意到的刺激会导致深层、全球递归处理,达到更高的水平,包括前额、执行区域,为大脑的执行系统提供可访问的内容(AC,除了PC),以用于工作记忆、报告和其他行为反应,即GNWT中的“点燃”。
在RPT中,注意力对于确定哪些刺激导致全球(阶段4)递归处理和AC,以及局部(阶段3)递归处理和PC至关重要。只有在注意力的作用下,主体才能“知道”并使用和报告其体验(AC)。RPT还声称,不同形式的短期记忆映射到PC和AC。全球RP参与稳健、低容量的工作记忆(WM),而局部RP则参与大容量、非常短暂的感官(图像)和脆弱记忆(SM,FM;0.1-几秒)在感官皮层中,这种记忆很容易被其他刺激覆盖。这解释了例如注意盲和变化盲现象,以及为什么在没有提示的情况下,只有少数视觉项目(保存在WM中)会在短暂展示后被报告。
RPT是由实验者开发的,扎根于实证观察,而不是更哲学的理论。大量实验结果支持RPT,例如来自遮蔽、变化盲、注意盲、注意闪烁、忽视和消失等现象,但并不排除其他理论。早期(<200毫秒)事件相关电位(ERP)成分(视觉意识负性;VAN)似乎与局部RP(阶段3)相匹配,并可能反映PC。晚期(约300毫秒)阳性(LP)ERP波(P300/P3b)显示出阶段4的特征,因此可能反映AC,与GNWT一致。
RPT面临的一个挑战是RP是否“足以产生意识”。尽管一些看似否定RPT的实证挑战通过区分PC和AC得以解决,但核心问题仍然存在:RP似乎比意识更为普遍,表明RP本身并不足以产生意识。实际上,RP在明显无意识状态(无梦睡眠、麻醉、昏迷)中并未完全消失。然而,Lamme建议“缺失的成分”可能是神经可塑性,因为RP可能满足Hebb的规则,从而“根本上与前馈处理不同”。因此,意识处理可能有助于实现感知学习。
预测处理与神经表征主义 预测处理(PP)及其更具体的形式神经表征主义(NREP)作为意识理论的关键思想是,大脑并不是通过提取或复制其周围世界的知识来完成感知,而是通过构建内部世界模型来建立感知和知识。这种构建通过学习对当前接收到的感官输入的最可能原因进行推断。因此,“预测”主要涉及对“此时此刻”的推断,而不一定是未来。通过比较内部表征的效果与实际感官输入,可以计算出预测误差,这为感知学习和推断服务。当特定情况在皮层-丘脑系统中实现时,推理表征可以变得足够广泛和丰富,以对应于意识表征,正如在感知、想象和做梦中发生的那样。
作为一种哲学立场,表征主义将感知对象和特征既不归属于物理的外部世界,也不归属于大脑本身(例如,如果我们意识到橙色,这种颜色并不真的在大脑中:神经元不会变成橙色)。NREP从这一立场出发,识别出产生PC的神经机制作为主要解释对象。NREP利用PP的计算原则构建基于感官的推理表征学习模型,假设运动行为并不是意识所必需的。在这里,NREP与基于PP的主动推断意识理论不同,后者认为运动行为(或“代理”)是意识的核心。
广泛的证据表明,预测或表征是实现大量感官、运动和认知过程的关键大脑功能,其中一些是有意识的,而另一些则不是。NREP中,意识现象的特征由五个不可剥夺的标志构成。这些标志共同定义了意识体验,作为主体世界的多模态、情境性调查,包括其身体。这种调查或“概述”(在多模态意义上)服务于但不等同于复杂决策和目标导向行为的规划。多个模态通过多感官整合的方式为这一空间广泛的调查做出贡献,不仅如此,还通过模态分离,使模态和子模态被体验为质上不同。在神经层面,这种调查由多个感官皮层层级(躯体感觉、听觉、视觉等)介导,这些层级各自具有前馈/反馈架构,并相互作用,以及与运动和规划系统的相互作用。在这个非常大的系统中,不仅需要单一模态中前馈和递归处理的相互作用,还需要模态之间的长程和侧向交互,以及在移动主体中,运动系统的交互。实际上,视觉皮层对感知刺激的递归波的主要部分可以解释为与运动相关的活动。因此,这种“超推断”的配置在其他能够生成预测的脑结构(例如小脑)中并不存在,因此这些结构不被认为与意识相关。NREP考虑了深度PP的几种神经实现,从严格的层级结构到异层级系统。
在这一观点下,意识和注意力是不同的过程:而多模态拓扑为基本意识体验提供了基础,注意系统则在该整体拓扑内突出有限的处理流。
树突整合理论 树突整合理论(DIT)源于对大脑皮层中递归处理作用的研究。它专注于皮层深层锥体神经元及其树突中非线性过程的作用,这使得这些神经元能够作为皮层-皮层和丘脑-皮层信息流的汇聚点。它(目前)基于六个核心观察,主要针对在啮齿动物中投射到亚皮层的(外皮层外,ET)5层锥体神经元: (1)5层锥体神经元通过非线性顶树突过程将到达不同树突区的信号关联起来,从而导致爆发放电。 (2)抑制这些顶树突过程(例如,树突钙峰)会抑制意识感知。 (3)抑制5层锥体神经元的亚皮层靶点(特别是高阶丘脑)会抑制意识感知。 (4)由于麻醉导致的意识丧失与这些神经元的顶树突与基树突之间的解耦相关。 (5)抑制高阶丘脑会导致顶树突与基树突的解耦。 (6)神经调节状态控制5层锥体神经元中的树突耦合。
综合这些发现,DIT认为,最好的解释是,投射到丘脑的亚皮层的顶树突轴作为皮层-皮层和丘脑-皮层信息流的汇聚点,同时是哺乳动物意识的基础。因此,DIT认为与这些5层神经元树突中与意识紧密相关的多种过程是理解其输出如何与意识紧密相关的基础。
DIT的基本见解是,5层皮层锥体神经元在两个不同区域接收和处理不同类别的信息:顶树突和基树突,这两个区域在物理和功能上是分开的。这两个树突区之间的电气分离已经相当显著,由于高“泄漏”电导率,使得顶树突和体细胞的计算基本上是相互隔离的。然而,现在已经确立,这些神经元可以通过高度非线性过程将到达两个区的信息关联起来,这些过程依赖于树突中的电压敏感离子通道,并受到特定靶向抑制和神经调节的控制,作用于树突树的子域。DIT首先假设,这些神经元的基树突接收与正在处理的认知或感官特征相关的前馈信息。其次,它提出,长程反馈信息主要作用于同一神经元的顶树突,为大脑其他区域(包括皮层其他区域)中发生的过程提供信息(或上下文)。
在解释意识的神经机制时,拥有多个区的神经元的需求是DIT的一个独特特征,这导致许多推论和预测。例如,DIT的一个推论是,先前的知识(例如,语义记忆)与感知过程的关联应依赖于连接到顶树突的突触的影响。最近在啮齿动物中获得的证据表明,巩固的语义记忆存储在连接于第1层的突触中,并影响顶树突的活动。此外,观察到高阶丘脑投射维持顶树突与基树突之间的耦合,已被提出为解释皮层-皮层和丘脑-皮层环路的同时完成的机制,因此可以作为(麻醉引起的)意识丧失的机制。这反过来又导致假设,注意力的定向(AC的一个方面)由不同皮层柱中树突耦合的选择性维持控制,这反过来可以调节与PC相关的后皮层区域中锥体神经元的耦合。抑制性和神经调节输入对耦合的调节也对大脑状态与意识之间的关系做出具体预测,包括做梦。
比较理论的聚合层次 在这里,我们比较所有五种理论,讨论它们的相似性和差异,以及兼容性和互补性。所有理论都提出了通常重叠或相似的神经机制,提供了潜在的聚合点。神经机制可以在不同的组织层次上进行描述,从微观的(亚)细胞过程,到中观的区域内电路,再到宏观的脑区域之间的相互作用。尽管不同的意识理论主要关注不同的层次,但这些层次被视为层次化组织,使得描述它们的机制关系并辨别差异和共性成为可能。
框3 如何推进意识理论:剩余差异、批评和方法 当前对神经意识理论的整合方法并不意味着应对剩余的根本差异和批评持天真的态度。除了对PC与AC的不同关注外,还有几个关键差异值得强调。尽管GNWT、RPT、PP/NREP和DIT强调信息处理、计算和消息传递在生成体验内容中的作用,但IIT将体验与神经基质支持的因果-效果结构相同。这要求考虑不仅是基质的实际状态,还要考虑其潜在状态的全貌,而其他理论则将特定的意识体验解释为实际状态,而不考虑系统的所有可能状态。因此,IIT并不将意识与特定功能等同(见Tononi和Koch),而其他理论则将意识的神经基础视为功能性方面,例如信息分布(GNWT)、学习和特征绑定(RPT)、预测(PP、DIT和NREP)以及多模态情境调查(NREP)。然而,这些理论及其他理论将功能方面归因于意识并不等同于计算功能主义(例如,Pennartz)。
对神经意识理论的一个普遍批评是,在考虑了图1-7中所示的机制后,人们仍然可能会问:“……但这些机制为什么会产生意识?”这个问题涉及到“解释差距”或“硬问题”,不同理论以不同方式处理这一问题(见正文),但也应被视为严肃的批评。例如,普遍接受的递归处理原则可能对意识是必要的,但不太可能是充分的。因此,当前的理论可能面临约束不足的问题,需要进一步的实证证据和理论推理来加以阐述。一个例子是,缺乏约束可能导致对意识在非神经基质中的差异性宽容(例如,IIT相对宽容;NREP和DIT则更为严格)。此外,除了“结构”和“消息”之间的区别外,如何使消息获得意义(意向性,“关于性”)的问题仍然在背景中徘徊(例如,Pennartz、Chalmers和Searle)。
如何在这个似乎令人畏惧的哲学雷区中前进?一种方法是对理论进行对抗性测试,针对它们在可实证预测上的差异。五个这样的项目已由Templeton世界慈善基金会赞助,包括对GNWT、IIT、PP/NREP和RPT的测试。实验目标,例如代表意识内容的解剖位置、刺激持续时间和点燃过程的作用等。然而,我们认为这种对抗性测试将受益于互补的方法,因为它可能(1)测试与正在审查的理论无关的预测,无法测试其核心假设,以及(2)过于依赖于反驳主义,其中期望对理论有明确的接受/拒绝二分法。一个理论下看似错误的东西,在另一个理论下可能是完全可接受的,理论的改进可能更依赖于对某些观点的逐步证据支持,符合溯因推理(“对数据的最佳解释推断”)和贝叶斯证据积累。因此,我们主张一种互补的方法,通过积累新的实证证据,结合各种理论中支持的最佳元素,形成一个改进的框架,具有更强的预测能力。最后,我们还期望这些结合的方法逐步推动一些看似深刻的概念(有时是形而上学的)分歧的回退。例如,定义更具实证基础的理论约束预计将限制可以合理归因于意识的复杂自然系统的范围。
微观层面:细胞和亚细胞层面 在意识科学的许多研究中,一个共同的假设是,意识可以在大规模网络及其计算层面上最好地理解。然而,在大脑中,全球活动依赖于电路、细胞甚至亚细胞层面上发生的过程。在这里,我们采取多尺度的方法,试图揭示一些与意识相关的过程,从细胞层面开始。
递归处理和整合——这两个在神经系统各个层面上运作的强大原则——在大多数意识理论中都是核心,包括DIT、RPT、GNWT、IIT和PP/NREP。DIT特别假设,递归和整合是通过5层锥体神经元中的树突非线性过程紧密相连的,这些神经元是丘脑-皮层环路信息流的关键枢纽。在清醒、意识状态下,皮层中的5层锥体细胞处于“耦合状态”,因此到达顶树突的反馈可以影响细胞层面的体细胞处理,从而将反馈中包含的内部信息重新整合到整个系统中,以大规模并行的方式进行。根据DIT,伴随意识状态的树突整合确保整个丘脑-皮层系统处于一个可以在空间和时间上整合活动的状态。这可能意味着活动可以在皮层-皮层递归连接中传播和扩散,锥体细胞可以激活高阶丘脑核,这反过来又可以增强皮层中的活动。随着活动的演变,产生了通过宏观尺度测量的意识的丰富全球动态。
DIT假设,这种局部机制为其他处理层次中与意识相关的过程奠定了基础。例如,IIT预测,锥体细胞的解耦会导致主复合体内因果-效果谱的急剧崩溃,从而降低Φ值,因此意识消失。类似地,如果树突整合是丘脑-皮层网络中信息流的核心,全球工作空间活动将在顶树突和基树突的神经元解耦时崩溃。因此,导致大规模点燃的丘脑-皮层系统中足够的活动传播和整合,可能仅在5层锥体细胞处于耦合状态时发生。类似地,局部树突整合可能使得在多个感官模态和跨尺度中建立联合神经活动模式成为可能,这在NREP中至关重要。总之,亚细胞树突整合可能是丘脑或皮层网络中大规模整合的一个必要条件。因此,DIT所提出的机制可以被纳入任何依赖于这种大规模整合的意识理论中。
树突整合还为具有双流架构的意识理论(如RPT和经典PP)提供了神经生物学的细微差别。特别是,关于树突整合机制的知识指定了两个流在单个锥体神经元层面上的交互方式。根据RPT,意识源于大脑中不同处理层次之间的递归相互作用。这个观点与DIT兼容,因为锥体神经元的解耦会导致大脑中RP的崩溃。根据与PP相关的意识理论(例如NREP),从高层到低层处理的预测试图“解释掉”从低层到高层的活动传播,通过抑制与预测相匹配的感官信号。无法预测的信号将作为预测误差传递到更高层次。尽管根据DIT,锥体神经元的内在生物物理过程规定了它们放大两个流之间的匹配,但有多种方式可以将锥体神经元或相邻皮层抑制神经元的生物物理特性纳入PP框架。所有这些变体都与树突整合的原则兼容,作为生成意识体验的细胞基础。然而,根据DIT,递归只是意识的一个必要条件。另一个关键组成部分是树突整合机制,使得将反馈信息纳入丘脑-皮层系统的持续激活模式成为可能。
中观层面:局部电路和递归处理 局部电路内和跨丘脑皮层微电路的处理如何可能对意识产生贡献?除了DIT中强调的5层神经元外,连接数百万突触的多种神经元类型在皮层内和跨皮层列中形成了极其复杂的微电路,这种复杂性很容易被低估。这种密集的复杂性,加上同样丰富的区域间和皮层-亚皮层相互作用,可能超越人类的直觉,并可能(不可预测且意外地)对主观体验(PC)产生或导致似乎非物理的现象。五种理论中讨论的局部、嵌套的复杂性可能与意识相关,并可能对意识产生贡献,例如与WM(图3)和感官适应相关的过程。然而,局部处理具体能为意识贡献什么?基于DIT的预测,PP/NREP提出,低层次预测和预测误差的计算在于调节信息流动的局部皮层电路,尤其是L5,但也包括L2/3的锥体细胞。越来越多的证据表明,表层锥体细胞在表征(预测)和误差编码中发挥作用,而L5细胞可能特别用于表征编码。特别是,阳性神经元在误差计算中被认为发挥作用。
尽管RPT主要关注视觉区域内和之间的RP,Lamme也讨论了即使在相互连接的神经元对之间的RP是否可能产生某种意识。然而,反馈的普遍性使Lamme提出,可能还需要其他过程,例如神经可塑性,才能产生PC。
GNWT暗示,意识访问(AC)依赖于RP,因为其“点燃”和回响活动实现了对意识内容的访问。早期版本的GNWT提出,L2/3的锥体细胞(而不是DIT中的L5神经元)对AC至关重要,但最近的版本也考虑了与亚皮层结构的环路。Dehaene等人的AC模型,包括丘脑皮层列的层级前馈和反馈投射,模拟了通过P300波和注意闪烁实现的点燃。网络模型提出,短暂(<1秒)的感知缓冲区可以被访问并广播到WM和意识(AC),通过同一电路的不同操作模式。根据IIT,反馈(RP)对意识是必要的。根据如何计算Φ以“量化”PC,具有纯前馈架构的人工系统的Φ=0,因此缺乏意识(PC),即使它们的可观察性能与具有意识的系统(Φ>0)相同(但见Doerig等人)。因此,即使在适当的架构存在的情况下,如果功能上禁用反馈连接,意识也应消失。另一方面,由于IIT认为脑电路的因果结构,而不是它们的活动,决定了Φ和PC,因此在主Φ复合体内因果整合的电路可以对PC做出贡献,即使其神经元处于静息状态。这一反直觉的预测并不被其他理论所共享,正在IIT与PP之间的对抗性合作中进行测试(见:www.templetonworldcharity.org/projects-database/0646;框3)。
Haun和Tononi提出,感官皮层中的“网格状”神经电路,类似于格子,形成了空间体验的基质。网格状连接在视觉和其他感官区域中均可发现,但PP相关的账户通常以不同的方式解释感官皮层的拓扑结构,而不是将其视为从感官(例如视觉)空间的一部分推断到另一部分的布局(参见Knierim和Van Essen)。此外,NREP提出,主观空间体验源于在不同空间框架(视网膜、颅内和外部)之间的多模态整合,这些框架由枕叶、颞叶和顶叶皮层使用。
除了表征信息内容外,局部电路还会随着不同“层次”的唤醒状态发生变化(反映出经典的“意识维度”之一);但请参见Hill和Tononi、Laureys等人以及Bayne等人对这一观点的更近期看法。然而,根据对神经调节状态的更近期观点,这些状态可以被视为“背景条件”或使能因素,仅与意识本身间接相关(未指定意识内容)。不同的局部“唤醒”状态和由麻醉或神经调节引起的网络复杂性变化已在丘脑-皮层网络模型中进行了建模,似乎与这里讨论的五种理论大致兼容,尽管DIT尚未包括这些。即使在孤立的皮层切片(“中观尺度”)中,这些调节效应也已通过修改的扰动复杂性指数(PCI)和与人类相似的复杂性指数进行评估(见框1;总体概念:丰富性和复杂性)。尽管这些测量无法推断皮层切片和啮齿动物的意识水平,但它们可以提供有关意识丧失和复杂性恢复的脑状态变化的细胞和网络机制的信息。
宏观层面:多区域系统 当沿着神经解剖学的尺度上升时,我们就会面临一个问题,即支持意识体验的脑机制如何在由多个区域组成的脑区域和系统层面上运作。这里讨论的五种理论可以分为三个部分。首先,IIT、PP-NREP和RPT共同认为,在意识视觉中,后皮层(更具体地说,涉及视觉的枕叶、顶叶和颞叶区域)发挥着最重要的作用。IIT认为,当高低视觉区域之间发生密集的再入交互时,最大Φ值将主要在后皮层(“热区”)达到。PP相关的账户,包括NREP,提出,意识视觉始于前馈信号流(主要传递误差信息)与递归流(传递自上而下的预测)之间的相互作用,尽管可能存在替代方案(例如,Heeger)。IIT和NREP都提出,在变化的视觉环境中,意识体验的不同、可切换的脑区域子集将参与前馈和反馈交互,以确保视觉特征的整合,具体取决于主体体验的视觉子模态。RPT在达到PC时也强调视觉皮层系统内的全球递归处理,而全球递归处理则需要注意力的作用,以便与前额区域的交互。
GNWT则赋予前额和顶叶皮层在意识视觉中的重要作用,这与其对AC而非PC的关注相一致,并将这种关注转化为与行为反应、注意力、WM和评估等相关的执行功能——所有这些功能都与前额-顶叶网络相关。第三,正如我们所看到的,DIT更强烈地关注意识的细胞基础。在这个意义上,DIT在“前部与后部”关于意识视觉的神经基质辩论中是“中立”的。值得注意的是,所有五种理论都一致认为,新皮层是生成意识内容的关键结构(与其他强调中心脑结构的理论相对),并且丘脑核在维持意识中至关重要。此外,所有五种理论都假设,除了局部或短程交互外,丘脑皮层系统中的长程通信也是意识所必需的。
然而,RPT和NREP都认为,视觉皮层层级内的信息处理不足以实现意识视觉(无论是理解为PC还是AC)。根据RPT,广泛的递归处理需要注意力,而为此,视觉皮层与前额区域之间的相互作用被认为是必要的。在这个意义上,RPT与GNWT比其不同的定义所暗示的更为兼容:它们都赋予前额-顶叶网络调节注意信号的关键作用。PP相关的账户将注意力解释为与大预测误差和低决策信心相关,促使主体寻求关于感官空间中感知不确定区域的更多信息。
在NREP定义的意识体验的标志中,我们发现动态性和稳定性的结合,其中“稳定性”指的是尽管发生眼睛、头部和身体运动,仍能将外部世界及其对象感知为稳定整体的能力。为了在身体活动的主体中实现视觉稳定,视觉皮层与运动皮层之间的相互作用是必需的:视觉系统接收运动信号(连同本体感觉和前庭再反馈)或关于正在发生或即将发生的感官变化的预测。此外,NREP假设,感官皮层系统之间的侧向、多感官交互对于确保意识内容的定性、多模态丰富性是必要的。因此,视觉不仅由经典定义的视觉丘脑皮层系统介导,而是由“扩展视觉皮层”介导,包括许多卫星区域。IIT和RPT在某种程度上并不与这些额外的交互相矛盾,但并未特别提及这一点。
因此,乍一看,IIT、RPT和NREP在意识视觉的宏观解剖学上似乎至少在整体上达成一致,而其他两种理论则要么表达不同的观点(与GNWT的不同),要么保持中立(DIT)。IIT、RPT和NREP的这种宏观层面的全球兼容性并不意味着它们具有相同的范围或相似性。例如,它们使用不同的理由来解释后皮层中递归反馈的必要性(IIT:最大化Φ;RPT:促进可塑性、学习和特征绑定;NREP:为表征学习和推断闭合循环)。另一个关键差异在于非活动神经元的假定作用(在IIT中,这些神经元可以对意识体验做出贡献,除非它们与网络断开或失活;其他理论则不持此观点)。尽管如此,我们可以得出结论,在多区域系统的层面上,这些理论并不像有时所假设的那样不兼容。
总体概念:丰富性与复杂性 越来越多的实证证据表明,与意识相关的神经交互模式不仅必须通过前馈和反馈连接紧密整合,还必须具有区分性和信息丰富性。这个脑复杂性的概念——定义为整合与区分的共存——在IIT最初的Φ概念中处于核心地位,并在其成熟形式的实证预测中仍然是中心。脑复杂性最近也被其他理论所接受。例如,GNWT的一个关键原则是广泛激活的点燃。因此,IIT和GNWT在实践中共享神经整合的基本要求,尽管它们的动机基础不同。值得注意的是,GNWT的原始表述并未明确提及与信息相关的测量、区分或复杂性,主要关注神经过程,例如前额-顶叶同步和P3b。随后,GNWT支持者的实证研究将重点转向动态谱的丰富性,最近,复杂性的概念已被明确纳入GNWT的关键元素,基于分布模块广播不同类型的感官信息。
根据IIT,脑复杂性与DIT和RPT所假设的微观和中观过程之间存在明确的机制关系。它要求有效的再入交互,以确保不同神经元组紧密结合并参与丰富的因果-效果交互。树突解耦和递归交互的崩溃预计将导致大脑整合信息的能力显著下降。沿着这些思路,在睡眠和无反应觉醒综合症(“植物人”状态)患者中观察到的复杂性崩溃与反馈交互的崩溃相关。
同样,复杂性是NREP所假设的多模态丰富性和整合的内在特征,尽管也存在重要差异。在IIT中,脑复杂性解释了每种体验的“直接”现象属性(统一性和丰富性),而NREP和PP则普遍将意识体验视为表征——对外部现实的最佳猜测“模拟”。在神经表征观点中,复杂性转化为一个多面现象,包括意识内容的多模态、定性丰富性、空间延展性和分辨率、强度和注意力,最后是时间深度——所有这些都被理解为有助于后续行动规划。因此,定义和操作化脑复杂性的方式有很多,未来需要更多的努力来明确基础假设并扩展复杂性的概念至一个详尽的分类法。
尽管如此,围绕复杂性的最近共识可能具有有趣的意义——尤其是在临床应用中,例如在无反应患者中进行更细致的分类(框1)。广泛的理论一致性可能增加这种信心,并为在更具挑战性的案例中推断提供合理依据。此外,复杂性和信息丰富性可以在多个层面上进行研究,从人类的外部和内部记录到啮齿动物和皮层切片。尽管在大脑中校准的当前复杂性测量不允许对意识水平进行推断,但它们提供了在不同尺度上建立神经过程之间的机制联系的前景,从微观尺度的树突整合到中观尺度的递归过程,再到宏观尺度的区域间整合。最后,这种围绕复杂性的共识可能代表了在更细粒度上对比不同框架的共同背景。
意识的复杂性在哪里以及如何在大脑中体现?对于IIT,具有递归金字塔状、汇聚-发散连接的网格状皮层结构(例如在后皮层中发现的)被认为对应于最大复杂性,因此足以支持意识体验。PP-NREP与IIT共享对空间延展感知的神经基础的追求,但在不同空间框架(视网膜、颅内和外部)之间的整合中寻找这一基础,这些框架由枕叶、颞叶和顶叶皮层使用;这些不同的方向可能证明是兼容的。另一个关键问题是:为什么一些意识理论关注正在进行的神经活动的复杂性,而其他理论则强调潜在因果结构的复杂性?这将是一个重要的澄清,因为这不仅涉及测量的不同方法(观察性与扰动性),而且还涉及对意识所需信息的根本不同理解(外在消息与内在因果-效果结构)。
总体评估:朝着整合的意识理论迈进? 调和理论是否现实,甚至渴望统一的意识理论?乍一看,正如我们所看到的,至少一些理论似乎是如此不同,以至于将它们桥接起来似乎不可逾越。这甚至导致了对竞争理论之间的合法性指控。然而,我们认为,多个理论的存在是这一新兴领域健康的标志,仔细关注解释对象可以理清核心主张,使得多个理论可以同时为我们的理解做出贡献。理论在关注和解释对象上的根本差异有时可能使它们更兼容。例如,GNWT专注于AC,而RPT、IIT和PP/NREP则主要试图解释PC。一个整合的意识理论必须在神经层面上解释PC,同时还要解释意识内容如何分布并用于后续的行动规划、WM、评估、注意调节等,这属于AC的领域。确实,朝着这一方向的建模努力正在进行,例如将PP账户与全球神经工作空间操作整合。我们还看到,RPT在原则上可以与GNWT整合,以共同努力包括AC。由于IIT已经承认前额区域在访问中的作用,因此与GNWT之间没有关于其“前部”AC的矛盾。然而,有人会争辩说,GNWT和IIT不能同时正确,因为它们在意识的本质上存在分歧。
潜在的概念聚合的进一步点在于高层次的复杂性相关概念,例如信息整合(IIT)、全球信息可用性(GNWT)、多模态丰富性中的多层次表征(NREP)、不同解剖尺度的递归处理(RPT)以及丘脑-皮层环路中的整合(DIT)。尽管表面上这些思想存在显著差异,但它们都基于不同方面的相互连接,承认神经交互的丰富性,依赖于丘脑-皮层系统中的短程和长程连接,是实现显著意识程度的重要因素。
这些聚合点和相关考虑为最终实现一个整合的意识理论——或者至少一组相互交织的概念和工作假设——提供了强有力的理由,可能在未来几年或几十年内变得可及。然而,也必须承认,仍然有广泛的理论和实证研究领域有待解锁。意识研究的现状可以与从不同角度检查的多面体物体进行比较,每个角度强调不同的方面,基于个体的视角。可以与量子力学进行类比,量子力学在观察和现象上达成一致,但在如何解释概率的微观描述与我们确定的宏观日常生活之间仍存在广泛的争议。这两种情况可能都反映了人类想象力和直觉的根本局限性。这一类比强调了意识研究人员开发明确概念和共同术语(例如,不仅仅使用“意识”这一通用术语来指代“AC”和“PC”等不同概念)以及将不同表述翻译为彼此的必要性。正如量子理论受益于强大的数学框架,这一领域也应如此。
最后,我们认为对意识的更多实验测试和实证测量的发展得到了广泛和无争议的支持。这一相关性有两个方面。首先,基于新理论的实证测量,如PCI和其他复杂性相关测量,可以在临床环境中有用,例如帮助将无反应患者分类为不同的亚组,进行更细致的区分和不同的恢复和治疗前景(例如,Demertzi等人、Casarotto等人和Farisco等人)。其次,实证测量可能有助于区分或统一不同理论。意识状态下的高复杂性时空EEG模式与递归激活波的存在始终相关,而低复杂性模式则在特征慢波睡眠和麻醉的皮层上下状态中发现。因此,PCI不仅与IIT的假设兼容,而且原则上也与GNWT、RPT、NREP和DIT相关的神经过程兼容。该领域已经受益于更广泛的实证工具和测量。DIT例如,引入了与不同意识阶段相关的EEG和LFP标记,PP/NREP则提出了依赖于解码主体所经历的意识内容的表征测量(参见Goltstein等人;见框2),而GNWT则有望从报告与无报告范式结合单元记录、fMRI和EEG-MEG中受益(例如,Sergent等人)。互补的计算方法可以用于促进进一步的比较和整合,研究正在从人类意识扩展到动物意识,以利用活体双光子成像、大规模集体记录、复杂性测量和光遗传学来揭示意识的因果基质。对抗性研究正在成为意识研究中的一种重要新方法,但最终是否会导致单一获胜理论的达尔文选择仍有待观察。在这里,我们主张一种更统一、整合导向的方法,试图实证测试并结合各种理论中的有价值元素。
结论 我们认为,神经意识理论(1)显示的差异往往并不像最初想象的那样大或不可逾越,(2) 提供了结合和整合不同层次组织中各种元素的机会,并且(3)一个理解大脑与意识关系的整合框架,在实证和理论进展的指导下,具有强大的潜力,能够在下一个十年推动该领域向前发展。
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.