DOI: 10.1126/science.282.5395.1846 , 1846 (1998); 282Science

et al.Giulio Tononi,Consciousness and Complexity

www.sciencemag.org (this information is current as of December 8, 2007 ):The following resources related to this article are available online at

http://www.sciencemag.org/cgi/content/full/282/5395/1846version of this article at:

including high-resolution figures, can be found in the onlineUpdated information and services,

http://www.sciencemag.org/cgi/content/full/282/5395/1846#otherarticles, 19 of which can be accessed for free: cites 61 articlesThis article

229 article(s) on the ISI Web of Science. cited byThis article has been

http://www.sciencemag.org/cgi/content/full/282/5395/1846#otherarticles 21 articles hosted by HighWire Press; see: cited byThis article has been

http://www.sciencemag.org/cgi/collection/neuroscienceNeuroscience

: subject collectionsThis article appears in the following

http://www.sciencemag.org/about/permissions.dtl in whole or in part can be found at: this article

permission to reproduce of this article or about obtaining reprintsInformation about obtaining

registered trademark of AAAS. is aScience1998 by the American Association for the Advancement of Science; all rights reserved. The title

CopyrightAmerican Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005. (print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last week in December, by theScience

on

Dec

embe

r 8,

200

7 w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

R E V I E W : N E U R O S C I E N C E

Consciousness and ComplexityGiulio Tononi* and Gerald M. Edelman

Conventional approaches to understanding consciousness are generally concernedwith the contribution of specific brain areas or groups of neurons. By contrast, itis considered here what kinds of neural processes can account for key propertiesof conscious experience. Applying measures of neural integration and complexity,together with an analysis of extensive neurological data, leads to a testableproposal—the dynamic core hypothesis—about the properties of the neuralsubstrate of consciousness.

W hat is the neural substrate of con-scious experience? While WilliamJames concluded that it was the

entire brain (1), recent approaches have at-tempted to narrow the focus: are there neu-rons endowed with a special location or in-trinsic property that are necessary and suffi-cient for conscious experience? Does primaryvisual cortex contribute to conscious experi-ence? Are brain areas that project directly toprefrontal cortex more relevant than thosethat do not (2)? Although heuristically useful,these approaches leave a fundamental prob-lem unresolved: How could the possession ofsome particular anatomical location or bio-chemical feature render some neurons soprivileged that their activity gives rise tosubjective experience? Conferring this prop-erty on neurons seems to constitute a catego-ry error, in the sense of ascribing to thingsproperties they cannot have (3).

Here, we pursue a different approach. In-stead of arguing whether a particular brainarea or group of neurons contributes to con-sciousness or not, our strategy is to charac-terize the kinds of neural processes that mightaccount for key properties of conscious ex-perience. We emphasize two properties: con-scious experience is integrated (each con-scious scene is unified) and at the same timeit is highly differentiated (within a short time,one can experience any of a huge number ofdifferent conscious states). We first considerneurobiological data indicating that neuralprocesses associated with conscious experi-ence are highly integrated and highly differ-entiated. We then provide tools for measuringintegration (called functional clustering) anddifferentiation (called neural complexity) thatare applicable to actual neural processes. Thisleads us to formulate operational criteria for

determining whether the activity of a groupof neurons contributes to conscious experi-ence. These criteria are incorporated into thedynamic core hypothesis, a testable proposalconcerning the neural substrate of consciousexperience (4).

General Properties of ConsciousExperienceConsciousness, as William James pointedout, is not a thing, but a process or stream thatis changing on a time scale of fractions ofseconds (1). As he emphasized, a fundamen-tal aspect of the stream of consciousness isthat it is highly unified or integrated.

Integration. Integration is a propertyshared by every conscious experience irre-spective of its specific content: Each con-scious state comprises a single “scene” thatcannot be decomposed into independent com-ponents (5). Integration is best appreciated byconsidering the impossibility of conceivingof a conscious scene that is not integrated,that is, one which is not experienced from asingle point of view. A striking demonstra-tion is given by split-brain patients perform-ing a spatial memory task in which two in-dependent sequences of visuospatial posi-tions were presented, one to the left and oneto the right hemisphere (6 ). In these patients,each hemisphere perceived a separate, simplevisual problem and the subjects were able tosolve the double task well. Normal subjectscould not treat the two independent visualsequences as independent, parallel tasks. In-stead, they combined the visual informationinto a single conscious scene and into a sin-gle, large problem that was much more dif-ficult to solve.

The unity of conscious experience is alsoevidenced by our inability to perform multi-ple tasks, unless some tasks are highly auto-matic and impinge less on consciousness.Moreover, we cannot make more than a sin-gle conscious decision within an interval of afew hundreds of milliseconds, the so-calledpsychological refractory period (7). Further-

more, we cannot be aware of two incongruentscenes at the same time, as indicated by thebistability of ambiguous figures and the phe-nomenon of perceptual rivalry (8). Unity alsoentails that conscious experience is private,that is, it is always experienced from a par-ticular point of view and cannot fully beshared (1).

Differentiation. While each consciousstate is an integrated whole, perhaps the mostremarkable property of conscious experienceis its extraordinary differentiation or com-plexity. The number of different consciousstates that can be accessed over a short time isexceedingly large. For example, even if wejust consider visual images, we can easilydiscriminate among innumerable sceneswithin a fraction of a second (9). More gen-erally, the occurrence of a given consciousstate implies an extremely rapid selectionamong a repertoire of possible consciousstates that is, in fact, as large as one’s expe-rience and imagination. Differentiationamong a repertoire of possibilities constitutesinformation, in the specific sense of reductionof uncertainty (10). Although this is oftentaken for granted, the occurrence of one par-ticular conscious state over billions of otherstherefore constitutes a correspondingly largeamount of information. Furthermore, it is in-formation that makes a difference, in that itmay lead to different consequences in termsof either thought or action.

The informativeness of consciousnesshelps dispose of many of the paradoxesraised about conscious experience. Consid-er a photodiode that can differentiate be-tween light and dark and then provide anaudible output, and a conscious human per-forming the same task and giving a verbalreport. Why should the differentiation be-tween light and dark performed by the hu-man be associated with conscious experi-ence, while presumably that performed bythe photodiode is not? The paradox disap-pears if one considers the information gen-erated by such discriminations. To the pho-todiode, the discrimination between dark-ness and light is the only one available, andis therefore minimally informative. To aconscious human, by contrast, an experi-ence of complete darkness and an experi-ence of complete light are two specificconscious experiences selected out of anenormous repertoire, and their selection im-plies the availability of a correspondinglylarge amount of information. To understandconsciousness, it is important to identify

The authors are at The Neurosciences Institute,10640 John J. Hopkins Drive, San Diego, CA 92121,USA.

*To whom correspondence should be addressed. E-mail: tononi@nsi.edu

S C I E N C E ’ S C O M P A S S ● R E V I E W

4 DECEMBER 1998 VOL 282 SCIENCE www.sciencemag.org1846

on

Dec

embe

r 8,

200

7 w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

underlying neural processes that are bothintegrated and capable of such exception-ally informative differentiations.

General Properties of NeuralProcesses Underlying ConsciousExperienceDistributed neural activity, particularly in thethalamocortical system, is almost certainlyessential for determining the contents of con-scious experience (4, 11). We suggested pre-viously that a key neural mechanism under-lying conscious experience are the reentrantinteractions between posterior thalamocorti-cal areas involved in perceptual categoriza-tion and anterior areas related to memory,value, and planning for action. Such interac-tions among neuronal groups in distributedbrain areas may be necessary in order togenerate a unified neural process correspond-ing to a multimodal conscious scene (4). Re-cent experimental findings are consistentwith this hypothesis and suggest some gen-eralizations about the neural processes thatunderlie conscious experience.

Activation and deactivation of distributedneuronal populations. Changes in specificaspects of conscious experience correlatewith changes in activity in specific brainareas, whether the experience is driven byexternal stimuli, by memory, or by imageryand dreams (12). Conscious experience assuch, however, involves the activation or de-activation of widely distributed brain areas(13), although what should count as the ap-propriate reference state for comparison isnot clear. In subjects who are comatose ordeeply anesthetized, unconsciousness is asso-ciated with a profound depression of neuralactivity in both the cerebral cortex and thal-amus (13). During slow-wave sleep, in whichconsciousness is severely reduced or lost,cerebral blood flow is globally reduced ascompared to both waking and REM (rapideye movement) sleep, two brain states asso-ciated with vivid conscious reports (14). Amore specific reference state would be theresponse to a simple sensory input when asubject is unaware of it versus when thesubject is aware of it. We have used magne-toencephalography to measure brain respons-es to flickering visual stimuli under condi-tions of binocular rivalry (15). A verticalgrating flickering at one frequency was pre-sented to one eye and a horizontal grating,flickering at a different frequency, was pre-sented to the other eye. Although the stimuliwere presented together, the subjects per-ceived either the vertical grating or the hori-zontal grating, with an alternation every fewseconds. It was found that the power ofsteady-state neuromagnetic responses at thefrequency of the flickering stimulus (its fre-quency tag) was higher by 30 to 60% in manysensor locations when the subject was con-

scious of that stimulus. The sensors withfrequency tags that correlated with consciousexperience were widely distributed over bothposterior (occipital and temporal) and anteri-or (frontal) areas. Furthermore, there wereconsiderable variations among different sub-jects (Fig. 1).

A change in the degree to which neuralactivity is distributed within the brain mayaccompany the transition between conscious,controlled performance and unconscious, au-tomated performance. When tasks are novel,brain activation related to the task is widelydistributed; when the task has become auto-matic, activation is more localized and may

shift to a different set of areas (16 ). In animalstudies, neural activity related to sensorystimuli can be recorded in many brain regionsbefore habituation. After habituation sets in(a time when humans report that stimuli tendto fade from consciousness), the same stimulievoke neural activity exclusively along theirspecific sensory pathways (17). These obser-vations suggest that when tasks are automaticand require little or no conscious control, thespread of signals that influence the perfor-mance of a task involves a more restrictedand dedicated set of circuits that become“functionally insulated.” This produces again in speed and precision, but a loss incontext-sensitivity, accessibility, and flexibil-ity (18).

Integration through strong and rapid re-entrant interactions. Activation and deactiva-tion of distributed neural populations in thethalamocortical system are not sufficientbases for conscious experience unless theactivity of the neuronal groups involved isintegrated rapidly and effectively. We havesuggested that such rapid integration isachieved through the process of reentry—theongoing, recursive, highly parallel signalingwithin and among brain areas. Large-scalecomputer simulations have shown that reen-try can achieve the rapid integration or “bind-ing” of distributed, functionally specializedneuronal groups dynamically, that is, in aunified neural process rather than in a singleplace (19, 20).

Substantial evidence indicates that the in-tegration of distributed neuronal populationsthrough reentrant interactions is required forconscious experience. An indication comesfrom the study of patients with disconnectionsyndromes, in which one or more brain areasare anatomically or functionally disconnectedfrom the rest of the brain due to some patho-logical process (21). In the paradigmatic dis-connection syndrome (the split brain), visualor somatosensory stimuli can activate thenondominant hemisphere and lead to behav-ioral responses, but the dominant, verbalhemisphere is not aware of them (22). Al-though the two hemispheres can still commu-nicate through indirect, subcortical routes,rapid and effective neural interactions medi-ated by direct reentrant connections are abol-ished by the lesion of the corpus callosum.Modeling studies suggest that a telltale signof effective reentrant interactions is the oc-currence of short-term temporal correlationsbetween the neuronal groups involved (19).Experiments on cats show that short-termtemporal correlations between the activity ofneuronal groups responding to the same stim-ulus, but located in different hemispheres, areabolished by callosal transections (23). Otherstudies indicate that various kinds of cogni-tive tasks are accompanied by the occurrenceof short-term temporal correlations among

Fig. 1. Amplitude and coherence differencesbetween the steady-state neuromagnetic re-sponses during binocular rivalry when subjectswere conscious of a stimulus and when theywere not. The differences are taken betweenamplitude and coherence values at 7.41 Hzwhen the subjects were conscious of a verticalgrating flickered at 7.41 Hz and when theywere not (that is, when they were conscious ofa horizontal grating flickered at 9.5 Hz). Ampli-tude differences are topographically displayedfor two subjects. Color scale is in picotesla.Significant positive differences in coherence at7.41 Hz between pairs of distant sensors areindicated by superimposed cyan lines. Bluelines indicate negative differences in coherence.Filled green circles indicate channels with sig-nal-to-noise ratio .5 that have coherence val-ues .0.3 with at least one other channel. See(15) for details.

S C I E N C E ’ S C O M P A S S

www.sciencemag.org SCIENCE VOL 282 4 DECEMBER 1998 1847

on

Dec

embe

r 8,

200

7 w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

distributed populations of neurons in thethalamocortical system (24). The magnetoen-cephalographic study of binocular rivalrymentioned above (15) also indicates thatawareness of a stimulus is associated withincreased coherence among distant brain re-gions (Fig. 1).

The requirement for fast, strong, and dis-tributed neural interactions may explain whystimuli that are feeble, degraded, or short-lasting, often fail to be consciously perceived.Although such stimuli may produce a behav-ioral response [perception without awareness(25, 26 )], they are unlikely to ignite neuralactivity of sufficient strength or duration tosupport fast distributed interactions. Con-versely, attention may increase the conscioussalience of certain stimuli by boosting thecorresponding neural responses as well as thestrength of neural interactions (27). Neuralactivity is also more likely to contribute ef-fectively to distributed neural interactions if itis sustained for hundreds of milliseconds.This would lead to the functional closure oflonger reentrant loops and thereby supportreentrant interactions among more distant re-gions (19, 20). Experimental findings areconsistent with this idea. High-frequency so-matosensory stimuli delivered to the thala-mus require about 500 ms for the productionof a conscious sensory experience, while lessthan 150 ms are sufficient for sensory detec-tion without awareness (28). The sustainedevoked potentials associated with a conscioussomatosensory sensation are apparently gen-erated by the excitation of pyramidal neuronsof primary somatosensory cortex through re-entrant interactions with higher cortical areas(29).

Evidence for a correlation between con-scious experience and sustained neural activ-ity also comes from tasks involving visuospa-tial working memory—the ability to rehearseor “keep in mind” a spatial location. Workingmemory is used to bring or keep some item inconsciousness or close to conscious access(30). In working memory tasks, sustainedneural activity is found in prefrontal cortex ofmonkeys, and it is apparently maintained byreentrant interactions between frontal and pa-rietal regions (31). Sustained neural activitymay facilitate the integration of the activity ofspatially segregated brain regions into a co-herent, multimodal neural process that is sta-ble enough to permit decision-making andplanning (32).

Differentiated patterns of activity. Al-though strong and fast reentrant interactionsamong distributed groups of neurons are nec-essary for conscious experience, in them-selves, they are still not sufficient. This isstrikingly demonstrated by the unconscious-ness accompanying generalized seizures andslow-wave sleep. During generalized sei-zures, the brain is not only extremely active,

but most neurons fire in a highly synchronousmanner. For example, the electroencephalo-gram (EEG) during petit mal absences indi-cates that groups of neurons over the wholebrain are either all firing together or all silenttogether, with these two neural states alter-nating every third of a second. Although suchhypersynchronous firing is indicative ofstrong and distributed interactions, a subjectwho is prey to such a seizure is unconscious.Similarly, during slow-wave sleep, neuronsin the thalamocortical system are active aswell as remarkably interactive, as shown bytheir synchronous firing in a stereotyped,burst-pause pattern. During this stage ofsleep, however, it is rare to obtain vivid andextensive conscious reports (33). By contrast,during REM sleep, when neural activity isnot globally synchronous but resembles therapid and complex patterns of waking, sub-jects typically report vivid dreams if awak-ened. We suggest that the low-voltage, fast-activity EEG characteristic of waking andREM sleep reflects the availability of a richand diverse repertoire of neural activity pat-terns. If the repertoire of differentiated neuralstates is large, consciousness is possible.Conversely, if this repertoire is reduced, aswhen most groups of neurons in the cortexdischarge synchronously and functional dis-criminations among them are obliterated,consciousness is curtailed or lost (34).

Theoretical Concepts and MeasuresThis brief review of neurological and neuro-physiological data indicates that the distrib-uted neural process underlying conscious ex-perience must be functionally integrated andat the same time highly differentiated. Asmentioned above, two key properties of con-scious experience are that it is integrated, inthe sense that it cannot be subdivided intoindependent components, and that it is ex-tremely differentiated, in the sense that it ispossible, within a short time, to select amongan enormous number of different consciousstates. It is a central claim of this article thatanalyzing the convergence between thesephenomenological and neural properties canyield valuable insights into the kinds of neu-ral processes that can account for the corre-sponding properties of conscious experience.Such an analysis requires the availability ofsatisfactory measures of integration and dif-ferentiation that can be applied to actual neu-ral processes, as well as an understanding ofthe neural mechanisms of integration.

Functional clustering: How to identify anintegrated process. How can one determinewhether a neural process is unified or simplya collection of independent or nearly inde-pendent subprocesses? We have suggestedthat a subset of distributed elements within asystem gives rise to a single, integrated pro-cess if, at a given time scale, these elements

interact much more strongly among them-selves than with the rest of the system — forexample, if they form a functional cluster.This criterion has been formalized by intro-ducing a direct measure of functional cluster-ing (35) which we summarize here.

Consider a jth subset of k elements (Xkj)

taken from an isolated neural system X, andits complement (X – Xk

j). Interactions be-tween the subset and the rest of the systemintroduce statistical dependence between thetwo. This is measured most generally by theirmutual information MI(Xk

j; X – Xkj) 5 H(Xk

j)1 H(X – Xk

j) – H(X), which captures theextent to which the entropy of Xk

j is account-ed for by the entropy of X – Xk

j and vice versa[H indicates statistical entropy (36 )]. Thestatistical dependence within a subset can bemeasured by a generalization of mutual in-formation, which is called integration and isgiven by I(Xk

j) 5 SH(xi) – H(Xkj), where

H(xi) is the entropy of each element xi con-sidered independently. We then define thefunctional cluster index CI(Xk

j) 5 I(Xkj)/

MI(Xkj; X – Xk

j) as a ratio of the statisticaldependence within the subset and the statis-tical dependence between that subset and therest of the system. Based on this definition, asubset of neural elements that has a CI valuemuch higher than 1 and does not itself con-tain any smaller subset with a higher CI valueconstitutes a functional cluster. This is a sin-gle, integrated neural process that cannot bedecomposed into independent or nearly inde-pendent components.

We have applied these measures of func-tional clustering both to simulated datasetsand to positron emission tomography dataobtained from schizophrenic subjects per-forming cognitive tasks (35). Theoreticallysound measures that can detect the occur-rence of functional clustering at the timescale (fractions of a second) crucial for con-scious experience may require additional as-sumptions. Nevertheless, it would appear thatthe rapid establishment of synchronous firingamong cortical regions and between cortexand thalamus should be considered as anindirect indicator of functional clustering,since it implies strong and fast neural inter-actions among the neural populations in-volved (19, 20). The mechanisms of rapidfunctional clustering among distributed pop-ulations of neurons in the thalamocorticalsystem have been studied with the help oflarge-scale simulations (19, 20). These haveshown that the emergence of high-frequencysynchronous firing in the thalamocortical sys-tem depends critically on the dynamics ofcorticothalamic and corticocortical reentrantcircuits and on the opening of voltage-depen-dent channels in the horizontal corticocorticalconnections (37).

Neural complexity: Measuring the differ-ences that make a difference. Once an inte-

S C I E N C E ’ S C O M P A S S

4 DECEMBER 1998 VOL 282 SCIENCE www.sciencemag.org1848

on

Dec

embe

r 8,

200

7 w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

grated neural process is identified, we need todetermine to what degree that process is dif-ferentiated. Does it give rise to a large reper-toire of different activity patterns or neuralstates? It is essential to consider only thosedifferences between activity patterns thatmake a difference to the system itself. A TVscreen may, for example, go through a largenumber of “activity patterns” that look differ-ent to an external observer, but that make nodifference to the TV.

A possible approach to measuring differ-ences that make a difference within an inte-grated neural system is to consider it as itsown “observer.” This can be achieved bydividing the system (which, we assume, con-stitutes a functional cluster) into two subsetsand then measuring their mutual information(38). The value of MI(Xk

j; X – Xkj) between

a jth subset Xkj of the isolated system X and

its complement X – Xkj will be high if two

conditions are met. Both Xkj and X – Xk

j musthave many states [their entropy must be rel-atively high (10)], and the states of Xk

j and ofX – Xk

j must be statistically dependent (theentropy of Xk

j must be largely accounted forby the interactions with X – Xk

j, and viceversa). The expression MI(Xk

j; X – Xkj) re-

flects how much, on average, changes in thestate of X – Xk

j make a difference to the stateof Xk

j, and vice versa.To obtain an overall measure of how dif-

ferentiated a system is, one can consider notjust a single subset of its constituent ele-ments, but all its possible subsets. The corre-sponding measure, called neural complexity,is given by CN(X ) 5 ½S ^MI(Xk

j; X – Xkj)&,

where the sum is taken over all k subset sizesand the average is taken over all jth combi-nations of k elements. Complexity is thus afunction of the average mutual informationbetween each subset and the rest of the sys-tem, and it reflects the number of states of asystem that result from interactions among itselements (39).

It can be shown that high values of com-plexity reflect the coexistence of a high de-gree of functional specialization and func-tional integration within a system, as appearsto be the case for systems such as the brain.For example, the dynamic behavior of a sim-ulated cortical area containing thousands ofspontaneously active neuronal groups (38)resembled the low-voltage fast-activity EEGof waking states and had high complexity.Such a system, whose connections were or-ganized according to the rules found in thecortex, visited a large repertoire of differentactivity patterns that were the result of inter-actions among its elements. If the density ofthe connections was reduced, the dynamicbehavior of the model resembled that of anoisy TV screen and had minimal complexi-ty. A large number of activity patterns werevisited, but they were merely the result of the

independent fluctuations of its elements. Ifthe connections within the cortical area wereinstead distributed at random, the systemyielded a hypersynchronous EEG that resem-bled the high-voltage waves of slow-wavesleep or of generalized epilepsy. The systemvisited a very limited repertoire of activitypatterns, and its complexity was low.

Measures of complexity, like measures offunctional clustering, can also be applied toneurophysiological data to evaluate the de-gree to which a neural process is both inte-grated and differentiated (40). This opens theway to comparisons of the values of neuralcomplexity in different cognitive and arousalstates and to empirical tests of the relation-ships between brain complexity and con-scious experience.

The Dynamic Core HypothesisA final issue we should consider is whether theneural process underlying conscious experienceextends to most of the brain, as was concludedby William James, or is restricted to varyingsubsets of neuronal groups. Several observa-tions support the latter possibility.

1) Classical lesion and stimulation studiessuggest that many brain structures outside thethalamocortical system have no direct influ-ence on conscious experience. Even withinthe thalamocortical system, many regions canbe lesioned or stimulated without producingdirect effects on conscious experience (41).

2) Neurophysiological studies indicate apossible dissociation between conscious ex-perience and ongoing neural activity withinportions of the thalamocortical system. Dur-ing binocular rivalry in monkeys, a largeproportion of neurons in early visual areas,such as V1, V4, and MT, continued to fire totheir preferred stimulus even when it was notconsciously perceived (42). The activity ofonly a subset of the neurons recorded in theseareas was correlated with the percept, al-though in higher areas such as IT and STS,the percentage reached 95%. In our magne-toencephalographic study of binocular rivalryin humans (Fig. 1) (15), we found that theresponses of only a subset of occipital, tem-poral, and frontal areas was correlated withthe conscious perception of a stimulus, al-though several other regions showed wide-spread responses to stimuli that were notconsciously perceived.

3) The firing of neurons dealing with rap-idly varying local details of a sensory input ora motor output does not seem to map toconscious experience. The latter deals withinvariant properties of objects that are highlyinformative as well as more stable and easilymanipulated. For example, patterns of neuralactivity in the retina and other early visualstructures correspond faithfully to spatial andtemporal details of the visual input and are inconstant flux. During each visual fixation,

however, humans extract the meaning of ascene and are not conscious of considerablechanges in its local details (43). Groups ofneurons responding in a stable way to invari-ant properties of objects are therefore morelikely to contribute to conscious experience.

4) Many neural processes devoted to car-rying out highly automated routines thatmake it possible to talk, listen, read, write,and so forth, in a fast and effortless way donot appear to contribute directly to consciousexperience, although they are essential in de-termining its content (44). As mentionedabove, neural circuits carrying out such high-ly practiced neural routines may becomefunctionally insulated except at the input oroutput stages. There is also some evidencethat cortical regions that are part of a fastsystem for controlling action, such as thedorsal visual stream, may not contribute sig-nificantly to conscious experience (45).

5) Although the sheer anatomical connec-tivity of the brain may hint that, over a suf-ficiently long time scale, everything can in-teract with everything else, modeling studiesindicate that only certain interactions withinthe thalamocortical system are fast and strongenough to lead to the formation of a largefunctional cluster within a few hundred mil-liseconds (46 ).

These observations suggest that changesin the firing of only certain distributed sub-sets of the neuronal groups that are activatedor deactivated in response to a given task areassociated with conscious experience. Whatis special about these subsets of neuronalgroups, and how can they be identified? Wesuggest the following:

1) A group of neurons can contributedirectly to conscious experience only if it ispart of a distributed functional cluster thatachieves high integration in hundreds ofmilliseconds.

2) To sustain conscious experience, it isessential that this functional cluster be highlydifferentiated, as indicated by high values ofcomplexity.

We propose that a large cluster of neuro-nal groups that together constitute, on a timescale of hundreds of milliseconds, a unifiedneural process of high complexity be termedthe “dynamic core,” in order to emphasizeboth its integration and its constantly chang-ing activity patterns. The dynamic core is afunctional cluster: its participating neuronalgroups are much more strongly interactiveamong themselves than with the rest of thebrain. The dynamic core must also have highcomplexity: its global activity patterns mustbe selected within less than a second out of avery large repertoire.

The dynamic core would typically includeposterior corticothalamic regions involved inperceptual categorization interacting reentrantlywith anterior regions involved in concept for-

S C I E N C E ’ S C O M P A S S

www.sciencemag.org SCIENCE VOL 282 4 DECEMBER 1998 1849

on

Dec

embe

r 8,

200

7 w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

mation, value-related memory, and planning(4), although it would not necessarily be re-stricted to the thalamocortical system. The term“dynamic core” deliberately does not refer to aunique, invariant set of brain areas (be theyprefrontal, extrastriate, or striate cortex), andthe core may change in composition over time(47). Because our hypothesis highlights the roleof the functional interactions among distributedgroups of neurons rather than their local prop-erties (2), the same group of neurons may attimes be part of the dynamic core and underlieconscious experience, while at other times itmay not be part of it and thus be involved inunconscious processes. Furthermore, since par-ticipation in the dynamic core depends on therapidly shifting functional connectivity amonggroups of neurons rather than on anatomicalproximity, the composition of the core can tran-scend traditional anatomical boundaries (48).Finally, as suggested by imaging studies (15),the exact composition of the core related toparticular conscious states is expected to varysignificantly across individuals.

The dynamic core hypothesis avoids thecategory error of assuming that certain local,intrinsic properties of neurons have, in somemysterious way, a privileged correlation withconsciousness. Instead, this hypothesis ac-counts for fundamental properties of con-scious experience by linking them to globalproperties of particular neural processes. Wehave seen that conscious experience is a pro-cess that is unified and private, that is ex-tremely differentiated, and that evolves on atime scale of hundreds of milliseconds. Thedynamic core is a process, since it is charac-terized in terms of time-varying neural inter-actions, not as a thing or a location. It isunified and private, because its integrationmust be high at the same time as its mutualinformation with what surrounds is low, thuscreating a functional boundary between whatis part of it and what is not. The requirementfor high complexity means that the dynamiccore must be highly differentiated—it mustbe able to select, based on its intrinsic inter-actions, among a large repertoire of differentactivity patterns. Finally, the selection amongintegrated states must be achieved withinhundreds of milliseconds, thus reflecting thetime course of conscious experience (49).

A number of experimental questions andassociated predictions are generated by thishypothesis. A central prediction is that, dur-ing cognitive activities involving conscious-ness, there should be evidence for a large butdistinct set of distributed neuronal groups thatinteract over fractions of a second much morestrongly among themselves than with the restof the brain. This prediction could, in princi-ple, be tested by recording, in parallel, mul-tiple neurons whose activity is correlatedwith conscious experience. Multielectrode re-cordings have already indicated that rapid

changes in the functional connectivity amongdistributed populations of neurons can occurindependently of firing rate (50). Recentstudies in monkey frontal cortex also showabrupt and simultaneous shifts among sta-tionary activity states involving several, butnot all recorded neurons (51). A convincingdemonstration of rapid functional clusteringamong distributed neuronal groups requires,however, that these studies be extended tolarger populations of neurons in several brainareas. Another possibility would be to exam-ine whether the effects of direct cortical mi-crostimulation spread more widely in thebrain if they are associated with consciousexperience than if they are not. In humans,the extent and boundaries of neural popula-tions exchanging coherent signals can beevaluated through methods of frequency tag-ging (15). Techniques offering both widespatial coverage and high temporal resolutioncould also help establish how large a dynamiccore normally is, how its composition chang-es, and whether certain brain regions are al-ways included or always excluded. It is alsosignificant to ask whether the dynamic corecan split, and thus whether multiple dynamiccores can coexist in a normal subject. Areasonable prediction would be that certaindisorders of consciousness, notably dissocia-tive disorders and schizophrenia, should bereflected in abnormalities of the dynamiccore and possibly result in the formation ofmultiple cores.

A strong prediction based on our hypoth-esis is that the complexity of the dynamiccore should correlate with the conscious stateof the subject. For example, we predict thatneural complexity should be much higherduring waking and REM sleep than duringthe deep stages of slow-wave sleep, and thatit should be extremely low during epilepticseizures despite the overall increase in brainactivity. We also predict that neural processesunderlying automatic behaviors, no matterhow sophisticated, should have lower com-plexity than neural processes underlying con-sciously controlled behaviors. Finally, a sys-tematic increase in the complexity of coher-ent neural processes is expected to accompa-ny cognitive development.

The outcome of such tests should indicatewhether conscious phenomenology can in-deed be related, as we suggest, to a distrib-uted neural process that is both highly inte-grated and highly differentiated. The evi-dence available so far supports the belief thata scientific explanation of consciousness isbecoming increasingly feasible (52).

References and Notes1. W. James, The Principles of Psychology (Holt, New

York, 1890).2. F. Crick and C. Koch, Cold Spring Harbor Symp.

Quant. Biol. 55, 953 (1990); Nature 375, 121 (1995);

S. Zeki and A. Bartels, Proc. R. Soc. London Ser. B 265,1583 (1998).

3. G. Ryle, The Concept of Mind (Hutchinson, London,1949).

4. G. M. Edelman, The Remembered Present (BasicBooks, New York, 1989); iiii and G. Tononi,Consciousness: How Matter Becomes Imagination(Basic Books, New York, in press); see also G. Tononiand G. M. Edelman, in Consciousness, H. Jasper et al.,Eds. (Plenum, New York, 1998). pp. 245–280.

5. A “conscious state” is meant here as an idealization,exemplified by viewing a rapid succession of slides.

6. J. D. Holtzman and M. S. Gazzaniga, Neuropsycholo-gia 23, 315 (1985).

7. H. Pashler, Psychol. Bull. 116, 220 (1994). The dura-tion of this interval is comparable with the durationof conscious states [A. L. Blumenthal, The Processof Cognition (Prentice-Hall, Englewood Cliffs, NJ,1977)].

8. F. Sengpiel, Curr. Biol. 7, R447 (1997).9. H. Intraub, J. Exp. Psychol. Hum. Percept. Perform. 7,

604 (1981); I. Biederman, Science 177, 77 (1972).10. C. E. Shannon and W. Weaver, The Mathematical

Theory of Communication (Univ. of Illinois Press, Ur-bana, IL, 1963). Note that the informativeness ofconsciousness also helps us to understand its evolu-tionary value (4).

11. V. B. Mountcastle, in The Mindful Brain, G. M. Edel-man and V. B. Mountcastle, Eds. (MIT Press, Cam-bridge, MA, 1978), p. 7; A. Damasio, Cognition 33, 25(1989); R. Llinas, U. Ribary, M. Joliot, X.-J. Wang, inTemporal Coding in the Brain, G. Buzsaki, R. Llinas, W.Singer, Eds. (Springer-Verlag, Berlin, 1994); J. New-man, Consciousness Cognit. 4, 172 (1995); T. W.Picton and D. T. Stuss, Curr. Biol. 4, 256 (1994).

12. R. S. J. Frackowiak, Human Brain Function (Academ-ic Press, San Diego, CA, 1997); P. E. Roland, BrainActivation ( Wiley-Liss, New York, 1993); M. I. Pos-ner and M. E. Raichle, Images of Mind (ScientificAmerican Library, New York, 1994). These imagingstudies confirm and extend previous lesion andstimulation studies.

13. Lesion studies indicate that consciousness is abol-ished by widely distributed damage but not by local-ized cortical damage. The only localized brain lesionsresulting in loss of consciousness typically affect thereticular core in the upper brainstem and hypothal-amus or its rostral extensions in the reticular andintralaminar thalamic nuclei [F. Plum, in Normal andAltered States of Function, A. Peters and E. G. Jones,Eds. (Plenum, New York, 1991), vol. 9, p. 359]. Al-though it has been suggested that the reticular coremay have a privileged connection to conscious expe-rience [ J. E. Bogen, Consciousness Cognit. 4, 52(1995)], its activity may simply be required to sustaindistributed activity patterns in the cortex.

14. A. R. Braun et al., Science 279, 91 (1998); P. Maquetet al., Nature 383, 163 (1996). Neural activity inslow-wave sleep is reduced in both anterior neocor-tical regions (most of the prefrontal cortex), as wellas in posterior cortical regions (especially parietalassociation areas), in paralimbic structures (anteriorcingulate cortex and anterior insula), and in centren-cephalic structures (reticular activating system, thal-amus, and basal ganglia); in contrast, it is not de-pressed in unimodal sensory areas (primary visual,auditory, and somatosensory cortex).

15. G. Tononi, R. Srinivasan, D. P. Russell, G. M. Edelman,Proc. Natl. Acad. Sci. U.S.A. 95, 3198 (1998); R.Srinivasan, D. P. Russell, G. M. Edelman, G. Tononi,Soc. Neurosci. Abstr. 24, 433 (1998).

16. S. E. Petersen, H. vanMier, J. A. Fiez, M. E. Raichle,Proc. Natl. Acad. Sci. U.S.A. 95, 853 (1998); R. J.Haier et al., Brain Res. 570, 134 (1992).

17. J. A. Horel et al., Science 158, 394 (1967).18. B. J. Baars, A Cognitive Theory of Consciousness (Cam-

bridge Univ. Press, New York, 1988).19. O. Sporns, G. Tononi, G. M. Edelman, Proc. Natl.

Acad. Sci. U.S.A. 88, 129 (1991); G. Tononi, O.Sporns, G. M. Edelman, Cereb. Cortex 2, 310 (1992).

20. E. D. Lumer, G. M. Edelman, G. Tononi, Cereb. Cortex7, 207 (1997); ibid., p. 228. For example, in a large-scale model of the visual system, reentrant interac-tions between groups of neurons in perceptual or“posterior” areas and in executive or “anterior” areas

S C I E N C E ’ S C O M P A S S

4 DECEMBER 1998 VOL 282 SCIENCE www.sciencemag.org1850

on

Dec

embe

r 8,

200

7 w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

rapidly led to their synchronous firing and to a cor-rect behavioral discrimination. This discriminationwas based on the dynamic binding of multiple visualattributes (position, movement, color, form) and ofdifferent levels of stimulus generalization (local fea-tures, invariant aspects of stimuli).

21. B. Kolb and I. Q. Whishaw, Fundamentals of HumanNeuropsychology (Freeman, New York, 1996). Psy-chiatric dissociation syndromes and conversion dis-orders may originate from a similar alteration ofreentrant interactions, although in these cases, thedisconnection would be functional rather than ana-tomical [ J. F. Kihlstrom, Consciousness Cognit. 1, 47(1992)]. Some explicit-implicit dissociations, such asamnesia, may also be due to a partial disconnectionof a lesioned area from the more global pattern ofneural activity that is associated with consciousness[D. L. Schacter, Proc. Natl. Acad. Sci. U.S.A. 89, 11113(1992)].

22. M. S. Gazzaniga, Neuron 14, 217 (1995).23. A. K. Engel, P. Konig, A. K. Kreiter, W. Singer, Science

252, 1177 (1991).24. S. L. Bressler, Brain Res. Rev. 20, 288 (1995); W.

Singer and C. M. Gray, Annu. Rev. Neurosci. 18, 555(1995); M. Joliot, U. Ribary, R. Llinas, Proc. Natl. Acad.Sci. U.S.A. 91, 11748 (1994); A. Gevins et al., Elec-troencephalogr. Clin. Neurophysiol. 98, 327 (1996).

25. A. J. Marcel, Cognit. Psychol. 15, 238 (1983); ibid., p.197; P. M. Merikle, Am. Psychol. 47, 792 (1992). Insome cases, perception without awareness has beenshown to occur with stimuli that are not short-lastingor weak [F. C. Kolb and J. Braun, Nature 377, 336(1995); S. He, H. S. Smallman, D. I. A. MacLeod, Invest.Ophthalmol. Visual Sci. 36, S438 (1995).

26. S. He, P. Cavanagh, J. Intriligator, Nature 383, 334(1996)].

27. J. H. Maunsell, Science 270, 764 (1995); K. J. Friston,Proc. Natl. Acad. Sci. U.S.A. 95, 796 (1998).

28. B. Libet, Ciba Found. Symp. 174, 123 (1993).29. L. Cauller, Behav. Brain Res. 71, 163 (1995).30. A. Baddeley, Proc. Natl. Acad. Sci. U.S.A. 93, 13468

(1996).31. J. M. Fuster, R. H. Bauer, J. P. Jervey, Brain Res. 330,

299 (1985); P. S. Goldman-Rakic and M. Chafee, Soc.Neurosci. Abstr. 20, 808 (1994).

32. The idea that neural activity must persist for a min-imum period of time in order to contribute to con-scious experience is also suggested by the phenom-enon of masking [ J. L. Taylor and D. I. McCloskey, Exp.Brain Res. 110, 62 (1996); K. J. Meador et al., Neu-rology 51, 721 (1998)].

33. M. Steriade, Cereb. Cortex 7, 583 (1997); D. Kahn,E. F. Pace-Schott, J. A. Hobson, Neuroscience 78, 13(1997).

34. Neural activity must also exhibit sufficient variancein time to support conscious perception. For example,if images on the retina are stabilized, perceptionfades rapidly, and a similar effect is seen in Ganzfeldstimulation. Short-lasting visual stimuli become in-visible if the transient neuronal responses associatedwith their onset and offset are suppressed by mask-ing stimuli [S. L. Macknik and M. S. Livingstone,Nature Neurosci. 1, 144 (1998)].

35. G. Tononi, A. R. McIntosh, D. P. Russell, G. M. Edel-man, Neuroimage 7, 133 (1998).

36. As a measure of statistical dependence, mutual in-formation has the virtue of being highly general,because it is multivariate and sensitive to high-ordermoments of statistical dependence [A. Papoulis,Probability, Random Variables, and Stochastic Pro-cesses (McGraw-Hill, New York, 1991)]. Note thatmutual information reflects a statistical dependenceamong subsets of a system, irrespective of its source.The presence and direction of causal interactionsbetween two subsets of a system can be evaluated,

at least in principle, by measuring the change inmutual information obtained by perturbing or de-efferenting each subset in turn.

37. These observations are of interest in view of thewell-known action of certain so-called dissociativeanesthetics, such as ketamine and phencyclidine, thatact as noncompetitive antagonists of the N-methyl-D-aspartate receptor [H. Flohr, Behav. Brain Res. 71,157 (1995)].

38. G. Tononi, O. Sporns, G. M. Edelman, Proc. Natl.Acad. Sci. U.S.A. 91, 5033 (1994). A complexitymeasure that does not involve the calculation ofaverage values of integration and mutual informationcan also be defined as the amount of the entropy ofa system that is accounted for by the interactionsamong its elements and is given by SMI(Xj

1; X – Xj1)

– I(X ) [G. Tononi, G. M. Edelman, O. Sporns, TrendsCognit. Sci., in press]. Note that complexity measuresshould be applied to a single system (a functionalcluster) and not to a collection of independent ornearly independent subsystems.

39. Changes in complexity can be obtained without mod-ifying the anatomical connectivity of the model bysimulating the transition between the burst-pausepattern of firing typical of slow-wave sleep and thetonic mode of firing typical of waking and REM sleep(G. Tononi, unpublished material). It should be notedthat high complexity is not easy to achieve. A systemof elements that are randomly interconnected, forinstance, may look very complicated, but it has lowvalues of complexity. On the other hand, systemsthat undergo selective processes so as to match thestatistical structure of a rich environment will grad-ually increase their complexity [G. Tononi, O. Sporns,G. M. Edelman, Proc. Natl. Acad. Sci. U.S.A. 93, 3422(1996)].

40. K. J. Friston, G. Tononi, O. Sporns, G. M. Edelman,Hum. Brain Mapp. 3, 302 (1995).

41. W. Penfield, The Excitable Cortex in Conscious Man(Thomas, Springfield, IL, 1958).

42. D. A. Leopold and N. K. Logothetis, Nature 379, 549(1996); D. L. Shenberg and N. K. Logothetis, Proc.Natl. Acad. Sci. U.S.A. 94, 3408 (1997). For otherinstances of dissociation, see (26); M. Gur and D. M.Snodderly, Vision Res. 37, 377 (1997); I. N. Pigarev,H. C. Nothdurft, S. Kastner, Neuroreport 8, 2557(1997); D. C. Bradley, G. C. Chang, R. A. Andersen,Nature 392, 714 (1998).

43. D. J. Simons and D. T. Levin, Trends Cogn. Sci. 1, 261(1997). The neurological evidence is in agreementwith these psychological observations. In the adult,lesions of the retina produce blindness, but they donot eliminate visual imagery, visual memories, andvisual dreams, while the latter are eliminated bylesions of certain visual cortical areas [M. Solms, TheNeuropsychology of Dreams (Erlbaum, Mahwah, NJ,1997)]. V1 may be important, however, to providevisual consciousness with a certain degree of detail.See also R. Jackendoff [Consciousness and the Com-putational Mind (MIT Press, Cambridge, MA, 1987)].

44. R. M. Shiffrin, in Scientific Approaches to Conscious-ness, J. D. Cohen and J. W. Schooler, Eds. (Erlbaum,Mahwah, NJ, 1997), p. 49; L. L. Jacoby, D. Ste-Marie,J. P. Toth, in Attention: Selection, Awareness, andControl, A. D. Baddeley and L. Weiskrantz, Eds. (Clar-endon, Oxford, 1993), p. 261; W. Schneider, M.Pimm-Smith, M. Worden, Curr. Opin. Neurobiol. 4,177 (1994).

45. A. D. Milner, Neuropsychologia 33, 1117 (1995);iiii and M. A. Goodale, The Visual Brain in Action(Oxford Univ. Press, New York, 1995).

46. The organization of the anatomical connectivity ofcertain brain regions, such as the thalamocorticalsystem, is much more effective in generating coher-ent dynamic states than that of other regions, such as

the cerebellum or the basal ganglia (G. Tononi, un-published material). Consistent with this, although incortical and thalamic areas 20 to 50% of all pairs ofneurons recorded are broadly synchronized, neuronsin the internal segment of the globus pallidus, theoutput station of the basal ganglia, are almost com-pletely uncorrelated [H. Bergman et al., Trends Neu-rosci. 21, 32 (1998)].

47. If the fast integration of neural activity comes at apremium in terms of number of connections andenergetic requirements, neuronal groups in “higher”areas should be privileged members of the dynamiccore underlying consciousness. Everything else beingequal, their firing is more informative, in the sensethat it rules out a larger number of possibilities. Forexample, the firing of face-selective neurons in areaIT considerably reduces uncertainty about a visualscene (seeing a face rules out countless other visualscenes), while the firing of retinal neurons reducesuncertainty by much less (a bright spot in a certainposition of the visual field is consistent with count-less visual scenes). The results of studies of binocularrivalry in monkeys and humans mentioned above areconsistent with this view.

48. We emphasize that the dynamic core, the highlycomplex, rapidly established functional cluster pro-posed to underlie conscious experience, is in no waythe only integrated but distributed neural processthat is relevant to brain function. We have hypoth-esized that distributed but integrated neural process-es called global mappings, encompassing portions ofthe thalamocortical system, as well as parallel loopsthrough cortical appendages such as the basal gan-glia, the hippocampus, and the cerebellum, underliethe unity of behavioral sequences (4). The functionalintegration of global mappings is envisioned to occurat longer time scales than the dynamic core (secondsas opposed to fractions of a second). However, thesetwo kinds of dynamic processes are expected topartially overlap for short periods of time.

49. Qualia—the seemingly inexplicable phenomenologi-cal manifestations of conscious experience—are con-ceived within this framework as rapid, highly infor-mative discriminations within a repertoire of billionsof neural states available to a unified neural processof great complexity. They correspond to the genera-tion of a large amount of information in a shortperiod of time. In this view, each quale—even aseemingly simple quale like a feeling of “redness”—corresponds to a discriminable state of the dynamiccore in its entirety, and not merely to the state of aspecific group of neurons in a certain brain area. Thesubjective meaning or quale of “redness,” for exam-ple, would be defined by the (increased) activity ofred-selective neurons as much as by the (reduced orunmodified) activity of neuronal groups selective forgreen or blue, for visual motion or shape, for auditoryor somatosensory events, and for proprioceptive in-puts, body schemas, emotions, intentions, and soforth, that jointly constitute the dynamic core. Thisview is antithetical to modular or atomistic ap-proaches to consciousness (2).

50. E. Vaadia et al., Nature 373, 515 (1995).51. E. Seidemann, I. Meilijson, M. Abeles, H. Bergman, E.

Vaadia, J. Neurosci. 16, 752 (1996).52. It is perhaps worth pointing out that our analysis

predicts the possibility of constructing a consciousartifact and outlines some key principles that shouldconstrain its construction. This work was carried outas part of the theoretical neurobiology program atThe Neurosciences Institute, which is supported byNeurosciences Research Foundation. The Foundationreceives major support for this program from Novar-tis Pharmaceuticals Corporation and the W. M. KeckFoundation.

S C I E N C E ’ S C O M P A S S

www.sciencemag.org SCIENCE VOL 282 4 DECEMBER 1998 1851

on

Dec

embe

r 8,

200

7 w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from