A NEURAL ATTENTIONAL MODELFOR ACCESS TO CONSCIOUSNESS
James Newman Colorado Neurological Institute
Denver, Colorado
and
Bernard J. Baars
The Wright Institute
Berkeley, California
from Concepts In NeuroScience, Vol. 4, No. 2 (1993) 255-290
(c) World Scientific Press
Send correspondence to: James Newman, Ph.D.
740 Clarkson Street
Denver, Colorado 80218
TEL: (303) 832-9062
EMAIL: newmanjb@aol.com
Key Words: attentional matrix; tangential intracortical network; pyramidal cells; nucleus reticularis thalami; binding
ABSTRACT
A broad consensus has developed in recent years in the cognitive and neurosciences that the cognitive functions of the mind arise out of the activities of an extensive and diverse array of specialized processors operating as a parallel, distributed system. A theoretical perspective is presented which expands upon this "society" model to include globally integrative influences upon this array of parallel processors. This perspective serves as the basis for an explicit neural model of a "global workspace within a system of distributed specialized processors". Anatomical and physiological evidence is reviewed which suggests that this parallel, modular architecture is superseded by a more diffuse, tangential intracortical network capable of integrating underlying modular activities into increasingly global cognitive representations. There follows an explication of the role of this "neural global workspace" in providing the essential basis for the central control of attention and the generation of unified, conscious percepts. Finally, the role of thalamic and brainstem activation systems in these integrative processes is discussed.
1.0 Introduction.
In recent years a consensus has been emerging in the cognitive and brain sciences which views the mind as being constituted of a vast "society" of specialized processors [1-5]. This view finds support in a considerable body of neuroscience evidence that characterizes the central nervous system as a parallel, distributed system of quasi-autonomous modular units [6,7]. It is widely believed that many, if not all, of the myriad cognitive functions of the mind/brain (from feature analysis to sentence parsing to mathematical computation, etc.), are performed by various coalitions of highly specialized sub-processors.
Missing from this scenario, however, are convincing answers to central questions regarding the nature of consciousness. How, for example, are the multiple representations generated by these multifarious processors combined to form what, subjectively, we experience as a stream of internally consistent perceptions, memories and thoughts? And how do the various activities taking place in these parallel circuits come under the control of voluntary attention? How do we employ our consciousness intentionally and intelligently to adapt to novel and constantly changing requirements?
Several competing theories have been put forward that address aspects of these questions [4, 5, 8-14]. In A Cognitive Theory of Consciousness, Baars [4] offers a framework for a comprehensive approach in terms of a single theoretical metaphor: a "global workspace (GW) system" underlying conscious experience. This framework attempts to encompass a number of contrasting perspectives on the nature of consciousness. These include the dichotomy between serial ("stream of consciousness") and parallel processing of information; between selective attention and diffuse cortical activation; and between conscious and unconscious cognitive processes. Baars [4] and Baars & Newman [15] pursue the hypothesis that these dichotomies reflect a certain large-scale architecture of the nervous system, one that is functionally equivalent to a global workspace in a system of distributed, specialized processors.
This theory for the conscious processing of information came, originally, out of efforts in Artificial Intelligence to more realistically model complex cognitive processes [16-18]. Computer researchers created a working memory, or general problem-solving space, where a large number of sub-processors are able to communicate their partial solutions to a general problem. In this problem-space coalitions of processors compete and cooperate with each other to evolve increasingly global solutions to complex cognitive tasks. The cooperating processors which come to dominate the workspace create a single, unified representation, or "global message", which is then broadcast back to all of the other processors, so that they can compete in successive, general solutions to ensuing problems. In this manner, problems which are novel to the system, or too complex to be solved using existing algorithms, can be addressed. This sort of architecture is alternately called a "blackboard" [18, 19], referring to the central representation of global messages so that they are accessible to an entire "audience" of specialized processors.
It is such a "neural GW" for the central representation of cognitive processes which Baars [4,8] hypothesized to form a necessary (though not sufficient) basis for conscious experience. This is in contrast to the great number of unconscious functions which the brain automatically carries out via changing arrays of parallel, distributed sub-processors. According to the theory, these parallel sub-processors: 1) are highly efficient in their specialized tasks; 2) operate automatically and relatively autonomously; 3) are highly diverse and; 4) together, possess immense processing capacity. Their functions are similar to those attributed by Kihlstrom [20] to "nonconscious mental structures and processes" that are "automatized through experience, and thus rendered unconscious." (p. 285).
In GW theory, these modular "cognitive units" are referred to as specialized processors. The term is not meant to designate a single, anatomically defined neural circuit. The existence of such "unit modules" or "cortical columns" in the cerebral cortex has been posited [6, 21], but there is presently no consensus among neuroscientists as to what constitutes a local cortical circuit. While "compartmentalization" of cortical circuitry is widely recognized, it can take a complex variety of forms [7, 11, 22]. In this paper, it is the principle of modular organization, not the specific circuitry, to which the term "specialized processor" refers.
The functions of these (generally) unconscious, specialized processors are very different from those attributed by GW theory to the conscious system. This system: 1) is highly serial in its operation; 2) has distinct limits on its information processing capacity; 3) is relatively slow and error-prone; but 4) possesses great range and integrative power in its representation of CNS activity [4, 8, 15]. These characteristics of the GW system are not modeled after any particular computer architecture. Rather they embody basic findings from cognitive research into the nature of conscious processes [4, 8].
In this light, this paper proffers a broad neural model developed out of Baars' [4] theoretical framework. This Neural Global Workspace (NGW) model views conscious processes in terms of a globally integrative brain system. The neural circuitry contributing to this system is not only widely distributed across the neocortex, but includes key corticothalamic and midbrain circuits as well. These cortico-subcortical circuits are hypothesized to be critical to understanding the mechanisms of attentional control that provide an essential basis for the conscious processing of information.
2.0 Basic Constraints and Assumptions of the Model.
The diversity of contemporary perspectives on attention and conscious awareness are such that some major constraints and assumptions of the present model need to be stated at the onset. In agreement with a long line of theorists from William James [23] up to the present day [11, 13, 24, 25], we assume that attention and consciousness are integrally related. Attention, however, is a poorly defined term, and encompasses numerous processes, operating at multiple levels of the nervous system [11, 26-29]. The present model is concerned with processes of attention related to the control of access to consciousness of mental contents. In the parlance of GW theory, this is equivalent to control of access to the global workspace of competing, potential global messages. In this context, the model limits its scope to attentional processes which are globally integrative and multi-modal, producing representations that selectively combine inputs from the major sensory modalities (and, potentially, any combination of specialized processors).
In the NGW model, global attentional systems mediate three broad states of conscious awareness: 1) orienting responses to novel or significant stimuli; 2) immediate perceptual awareness; and 3) focal awareness involving goal-directed behaviors. In developing the model, we will present evidence that these three attentional states have their basis in the activities of three interacting, global attentional systems: 1) the midbrain reticular formation; 2) the association areas of the posterior cortex; and 3) the prefrontal cortex. These three systems form an "attentional matrix" [30] centered upon the thalamus. Their projections upon the thalamus are highly ordered and allow them to regulate the flow of information reaching the cerebral cortex in highly differentiated ways. In this context, "global" does not mean vague or diffuse. Rather, it refers to a level of cognitive processing at which a single, coherent flow of information emerges out of the diverse activities of the CNS.
It should be emphasized that, in the present model, it is not the partial representations generated by individual processors which dominate in the workspace, but a global representation generated by their integrated activities. It is the global attentional systems which bind together such transiently dominant coalitions into unified conscious percepts. It is not the activity of these attentional systems of which we are subjectively aware, but the coherent stream of images they bind together. The entire process is highly recursive [10, 11] in that the percepts generated by these systems can then influence the subsequent activity of the attentional matrix.
3.0 The Binding Problem.
It follows from these assumptions that an adequate explanation of how consciousness arises in the nervous system must address the "binding problem" [13]. If consciousness involves the generation of a stream of highly integrated and coherent representations, how is this accomplished? We would note, parenthetically, that some prominent theorists [5, 31] maintain the binding problem is not a scientific problem at all, but a deeply rooted misperception of the phenomenological nature of awareness. And this view gains support from contemporary characterizations of the cognitive and neural apparatus as highly modularized.
Even a theorist like Edelman [11], who argues for an integrative, biological theory of consciousness, concludes that decades of research on visual systems of the brain has revealed "no single dominant, integrative high-level area ...; rather, every visual cortical area is connected to some subset of the other visual areas". [p. 70] Recent research studying the classical "association" areas further supports the view that no central, integrative area binds together inputs to the cortex. Goldman-Rakic [7] concludes,
the most recent data do not fully support the classical concept of association cortical areas as zones of increasing intersensory convergence. In fact, the large association areas, at least in non-human primates, can be divided into smaller specialized information centers that retain a large measure of modal specificity, at least as interpreted from anatomical considerations. (p. 182)
Thus, neuroscientific evidence seems to mitigate against "binding phenomena" having their origins in the workings of these parallel, distributed cortical circuits. In this paper, we will argue for an alternative hypothesis: that binding involves the imposition of a secondary process (or processes) upon the activities of these "smaller specialized information centers". Crick and Koch [13] have theorized about possible binding processes of this sort in the visual cortex. They theorize that "an attentional mechanism transiently binds together all of those neurons whose activity relates to the relevant features of a single vis-ual object .... by generating coherent semi-synchronous oscillations, probably in the 40-70 Hz range". [p. 263] Such a scenario was first proposed by von der Malsburg and Schneider [32].
Crick and Koch [13, 14] cite experimental work from several researchers over the past few years indicating that oscillations in this frequency range play an important role in perceptual processing in animals (although they acknowledge that the "40-Hz hypothesis" is far from being firmly established experimentally). In an independent review of animal experiments relating to the hypothesis, Stryker [33] concludes that "neurons in the visual cortex activated by the same object in the world tend to discharge rhythmically and in unison". [p. 297]
The focus of this paper is not on particular neural mechanisms but global, integrative systems. In consonance with hypotheses like Crick and Koch's, however, we will present a variety of evidence for the centrality of the correlated activation of widely separated ensembles of neurons in the cerebral cortex to the solution of the binding problem. We are hopeful in this regard that the prediction with which Stryker [33] ends his review will prove prescient.
Exploring the rhythms of the brain revered by the pioneers of electroencephalography but now mostly dismissed as irrelevant to neural information processing, may come back into fashion. [p. 298]
In this paper, we pursue the argument that "rhythmic cortical activation" and "information processing" are not so much incompatible, as complementary, ways of understanding neurocognitive processes. For this synthesis to make sense, however, earlier EEG explorations must be updated and translated into the language of present-day information theory. The model presented below offers an approach for integrating these two perspectives.
4.0 Traditional Candidates for Globally Integrative Substrates.
The basic premise of the Neural Global Workspace model is that conscious experience arises out of the global integration of multimodal CNS activity. Not only is our subjective experience one in which multiple perceptual and ideational factors are encompassed within a continuous series of unified mental representations, but that series of coherent images also serves as the essential basis for our conscious actions in the world. Given these prerequisites, a natural approach to searching for the neural basis of conscious experience is to identify those structures and functions which serve to globally integrate the brain's activities. The search for such "higher-level, integrative systems", however, has proved a fairly perplexing challenge for neuroscience.
Traditional candidates have been the "association" areas of the cortex, in particular the posterior parietal, inferotemporal and prefrontal regions, with their extensive and reciprocal connections. As already noted, however, contemporary knowledge of the anatomy and physiology of the cortex is more suggestive of a highly compartmentalized, than integrative architecture. More recently, attention has come to be focused upon the polymodal nuclei of the thalamus, with which association cortex has extensive reciprocal connections. Candidates which have been suggested as likely substrates for attention and consciousness include the pulvinar, ventral anterior, ventral medial, reticular and intralaminar nuclei [11, 14, 25, 34]. Excepting the pulvinar, these nuclei are part of (or closely connected with) what has traditionally been referred to as the "diffuse" or "nonspecific thalamus" [35, 36]. Herkenham [37] and Goldman-Rakic [7] argue, based upon anatomical studies of the laminar distribution patterns of thalamic projections in neocortex, that portions of several thalamic nuclei adjacent to the nonspecific thalamus are polymodal as well. Edelman's [11] model "relating attention and consciousness" includes basal ganglia and brainstem nuclei with which the nonspecific thalamus has extensive connections [p. 202]. Finally, Crick and Koch [14] have suggested the claustrum, a separate subcortical structure, may be involved in the integration of visual inputs to the cortex.
The cerebral cortex and thalamus are, in a very real sense, "mirror images" of each other. Each sends the bulk of its extrinsic projections to the other. The dozens of discrete nuclei that make up the thalamus send orderly sets of projections to nearly all of the primary and association areas of the cortex. Each cortical area, in turn, sends efferents back to the thalamic area from which the afferent projections are received, creating a dense fan of millions of reciprocal fibers extending from the prefrontal cortex to the occipital pole. As a result of this reciprocal connectivity, the gross anatomy of the neocortex is essentially recapitulated within the thalamus [30, 36].
Thalamocortical connectivity patterns have traditionally been characterized as either "specific" or "nonspecific". The specific nuclei of the thalamus (or principal nuclei, when those projecting to the cortical association areas are included) have traditionally been described as having well-delineated projections with the cortex. Each nucleus is distinct, having little interconnectivity with other nuclei [7, 30, 36]. The most specific of the principal nuclei are those that relay raw sensory data from the major sensory modalities (visual, auditory, and somatosensory) to the cortex. These sensory relay nuclei, like the primary projection areas with which they are connected, account for a relatively minuscule amount of the information processing capacities of the CNS. Despite this fact, most research into the neurophysiology of perception, from the 19th Century onward, has focused upon these highly specific systems.
4.1 The Nonspecific Thalamus.
During the 1950's and 60's, the nonspecific thalamus became the focus of intensive study as the diencephalic extension of the brain stem reticular activating system [30, 38]. Unlike specific nuclei, the nonspecific nuclei are extensively interconnected [30, 36]. Most of them lie within (or adjacent to) the internal medullary lamina, which forms a band of tissue separating the ventrolateral and mediodorsal thalamus. For the sake of simplicity, we will collectively refer to these nuclei as the "intralaminar complex" (ILC). The nucleus reticularis (NR) joins with the intralaminar complex at the anterior, or rostral, pole of the thalamus where both NR and ILC are considerably thicker. As NR and ILC extend posteriorly they form into two narrow sheets of tissue enclosing the ventral nuclei of the thalamus. Figure 1 illustrates this anatomy in a cross-sectional view of the interior of the thalamus. Also represented in the figure, in highly schematic form, are major patterns of corticothalamic connectivity.
------------------------------------------------------------------------------------
http://www.phil.vt.edu/ASSC/newman/fig7.html
-------------------------------------------------------------------------------------
As the figure shows, the nucleus reticularis surrounds the thalamus proper much like a shell surrounds an egg, or two eggs, there are actually a left and right thalamus, abutting at the midline. This bilateral symmetry is also illustrated in the figure. The nucleus reticularis is unique among all the thalamic nuclei in two respects: 1) it sends projections to all of the principal nuclei; and 2) it sends no projections to the cortex. It also has extensive interconnections with the brain stem reticular activating system, and receives collateral projections from nearly all of the neocortex [36, 37, 39, 40].
The intralaminar complex is truly multimodal, sharing extensive connections not only with NR, but the brain stem reticular formation and basal ganglia. The ILC is also the source of thalamocortical projections forming a "final common pathway" for the global activation of the cortex by the ascending reticular activating system [36, 37]. ILC projections are distributed across a number of cortical regions, but are most dense in motor and association cortex. Physiological research [35, 36, 41] has shown that stimulation of the nonspecific thalamus results in the diffuse, bilateral activation of wide areas of the cortex. The physiological distinction between the localized activation of the cortex produced by stimulation of specific nuclei, and the widespread activation associated with stimulation of the nonspecific thalamus (and reticular formation) was the initial basis for the specific versus nonspecific distinction [36, 37]. Collectively these nonspecific structures are sometimes referred to as the "reticular core".
As this brief review would suggest, the nonspecific thalamus, as part of the ascending reticular activating system, should be a prime candidate for a global, integrative substrate for consciousness. And indeed, this was an assumption of many of the neuroscientists who studied this system in the 1950s and 60s. That assumption had fallen into disfavor by the 1970s. In a review of the topic, Mesulam [30] concludes that predominantly subcortical theories of attention and consciousness have been all but supplanted by the consensus that "more complex functions are executed predominantly by neocortical mechanism." [p.134]
A major factor in this turning away from the concept of an ascending, subcortical activation system to cortically-centered theories of attention and consciousness has been that, for many neuroanatomists, the use of the terms nonspecific and diffuse caused controversy almost from the moment the distinction was coined. What seemed good descriptive terms physiologically, proved problematic for the anatomist. In studying the reticular core of the brain, it was discovered that while its connections were polymodal, and more diffusely organized than the classical sensory and motor tracts that surround it, reticular pathways show many indications of fairly orderly connections [41-43]. More recent anatomical research has shown that the projection patterns of the "diffuse" thalamus upon the cortex are not so diffuse as had been assumed, but resemble those of the parallel distributed systems of the neocortex. Anatomical studies of the distribution of projections of the intralaminar nuclei by Macchi and Bentivoglio [44] indicated that:
1) they establish reciprocal relationships with preferential regions of the cerebral cortex and
2) their wide cortical distribution is not sustained by a rich network of axon collaterals. [p. 389]
Thus, what appears from a physiological perspective to be a highly diffuse activating system has been shown, anatomically, to have a specifiable, parallel-distributed architecture. As if this were not enough to complicate the issue, anatomists have recently found "cell clusters" in the more medial portions of some of the principal thalamic nuclei that send out highly divergent pro-jections capable of innervating several cortical areas simul-taneously. These findings led Goldman-Rakic [7] to conclude that "the rigid distinctions between specific and nonspecific thalamic nuclei may no longer be appropriate." [p. 183].
What to make of these increasingly diverse permutations of connectivity? Do they point towards greater integration or greater specialization of function? Certainly integration must occur at many levels of a nervous system that is so densely
interconnected, but where are the globally integrative systems of the brain to be found? And how does a single, coherent stream of consciousness arise out of the diverse activities of a nervous system, which seems to have no central focus, let alone a central executive [6]?
This problem can be approached from a number of perspectives. Baars and Newman [15] builds upon the work of recent neuroscientists concerning the role of "reentrant circuits" distributed throughout the brain [11, 12]. The present model discusses several highly reentrant systems, but places its primary emphasis upon the role of global patterns of rhythmic activation in integrating cortical processing. This perspective has its basis in discoveries spanning fifty years of neuroscience research. To develop it properly requires a continuation of our brief historical survey.
5.0 Global Representation of Information in the CNS.
A pervasive quality of electrical activity in the brain is its rhythmicity. When EEG rhythms first began to be studied during the 1930s, it was widely believed that they represented the aggregate activity, or "envelope", of millions of action potentials. Subsequent research has demonstrated that rhythmic electrical potentials persist "even in the absence of propagated nerve impulses" [45, p. 11]. An early demonstration of this principle was the discovery by Green et al. [46], made during animal experiments in the late 1950s, that the theta rhythm (2-8 Hz) of the EEG associated with hippocampal activity arises, not from vertically transmitted action potentials, but,
from longitudinal currents flowing between the cell bodies of the pyramidal cells and the proximal two-thirds of the apical dendrites [of surrounding pyramidal cells]. The theta activity is not propagated all-or-none in this axis, nor do the waves represent the "envelopes" of action potentials, but may arise from non-propagated electrotonic changes, and synaptic potentials in the region of the apical dendrites, or both. [47, p. 520]
Such "region[s] of apical dendrites" are found across wide areas of the CNS. They generally occupy spaces where low densities of neuron cell bodies exist. The great bulk of synaptic junctions are found in such regions. These extra-cellular regions are filled with dendrites and axon collaterals branching off the main axons of cells. Taken together, these synaptic spaces form extended meshworks of dendrites and axon collaterals referred to by neuroanatomists as neuropil.
Adey et al. [47] postulated, based upon Green et al. [46] and their own research, that "the integration of information at the cellular level may occur on the basis of ubiquitous and pervasive wave processes" generated in the neuropil [p. 520]. Pribram [45] carried this argument a step further, arguing that information is only fully coded in the brain when these waves generated in the neuropil create an integrated "neural state".
The unit of analysis for brain function has classically been the neuron. The present proposal for a two-process mechanism recognizes an additional unit: the neural junction .... Neural junctions are thus much more than just way stations in the transmission of nerve impulses; they compose, at any moment, a neural state that is operated upon by arriving nerve impulses. In turn, nerve impulses generated by neurons are influenced by this state. (p. 25)
Schmitt et al. [48] reviewed evidence that these electrotonic currents have subtle, but pervasive effects upon synaptic transmission. As wave phenomena, they can spread these effects across extensive expanses of neuropil. But since these currents are localized to the synaptic junctions of neurons, they can also have highly specific effects upon the transmission of neural impulses. The authors contrast this electrotonic processing of information with that performed by the, many small computational units or modules providing for highly localized as well as rapid, parallel processing of input. In contrast, the tight elaborate weave of fine neuronal processes of the central neuropil at various sites constitutes a logical mechanism for the integration of ... activity of individual modules by ... connectivity within vast networks. The enormous complexity of the electrical and chemical interactions possible in such systems suggest that operation and output characteristics may not be predictable on the basis of unit properties alone; certain emergent properties related to field interaction may also be important. [48, p. 119]
Of the various sites where such "central neuropil" is found, by far the most extensive lies within the neocortex. Superimposed upon the thousands of vertically-oriented modular units, or "cortical columns" [6], arranged across the expanse of the neocortex is just such an "elaborate weave of fine neuronal processes" Szentagothai [21] was one of the earlier neuroscientists to contrast the anatomy, and possible functions, of this horizontally oriented, or "tangential", network with that of the vertically arranged cortical columns. He contrasted the anatomical organization of this tangential system with the columnar modules of the cortex.
the quasi-discrete modular spaces of the cortex ... are superseded by a much more diffuse system of connections that operate in spaces that are at least one order of magnitude larger. [p. 135]
The connections Szentagothai refers to are made by axon collaterals which branch off of the main axon of each pyramidal cell as it projects through the cortical lamina towards a distant efferent site. These axon collaterals radiate out tangentially, making synaptic connections with dozens of surrounding cortical columns (see Figure 2). Szentagothai [21] noted that,
------------------------------------
Figure 2 is not available here
------------------------------------
This system of connectivity is still local, but it is so wide that it has to merge with the relatively short-distanced cortico-cortical associative fibers .... the synapses [are] ... widely distributed, and in the upper cortical layers are engaged mainly with dendritic spines, probably of pyramidal cells. Additionally, there is a system of long surface parallel terminal axons in lamina I and the stria Gennari [layers III-IV] .... The pyramid cell collateral system and a large part of the lamina I system is obviously excitatory. [p. 135-136]
Kisvarday et al. [49], in a recent study of HRP-filled layer III pryramidal cells in the cat striate cortex, have confirmed that the synaptic targets of pyramidal axon collaterals are predominately the dendritic spines of other pyramidal cells, and that these connections are largely excitatory. Their research has also shown that these fibers spread tangentially in the cortical layers "often for several millimeters". They cite research going back to Ramon y Cajal in the late 1890s attesting to "The abundance of axon collaterals [as] one of the most impressive and most consistent features of the neocortical circuitry." [p. 541] Finally, they conclude that,
pyramidal cell axon collaterals may be responsible for the correlated firing of cell groups with similar orientation preferences that has been demonstrated to occur over distances of several millimeters in the cat cortex (Ts'o et al., 1986). [p. 549]
Szentagothai [21] describes the dynamics of this tangential system in the following manner,
The remarkable feature in the connections ... is that this system is completely (and circularly) symmetric and reciprocal in the surface plane: i.e. virtually the same connections exist from A to B as from B to A .... In such a system the conditions for cooperativity and the emergence of dynamic patterns are given [not only at the microscopic, but] also at the macroscopic level of the neuron network.
"Dynamic patterns" are most attractive because they offer "superstructures", developing over larger volumes of neural tissue, that are qualitatively different from, and at higher hierarchical levels than, the physiological operations conventionally known and studied by the neurophysiologist, and, what is particularly important, there is no difficulty to envisage such "dynamic patterns" as having their repercussions upon specific local operations of the neural network. Yet, they are phenomena that stay entirely within the framework of our natural laws. [p. 137]
5.1 Two Systems of Cortical Connectivity.
The nature of these dynamic patterns - and their "repercussions upon specific local operations" - is only beginning to be explored by neuroscience. In the section to follow, we will present the findings of two researchers working in this area [50]. At this point, however, we would like to offer a very simplified picture of how the dynamics of this tangential intracortical network might account for certain critical aspects of Global Workspace theory. To summarize what has been discussed up to now concerning the circuitry of the neocortex: it can be conceptualized in terms of two uniquely different, but highly interconnected, systems. The first consists of quasi-discrete columns of neurons with very specific, vertically-oriented patterns of connectivity. The second system spreads tangentially, in a highly diffuse and continuous pattern of connectivity, across the modular spaces of the cortex. A remarkable feature of this neocortical architecture is that at the heart of both the vertical and tangential systems are the pyramidal cells of the cortex.
Pyramidal cells, of course, are primarily responsible for the long distance transmission of neural impulses. Their axons not only carry pulse-coded signals to and from most areas of the cortex, but to subcortical structures as well, including those involved in the initiation of motor responses. When a pyramidal cell fires, two things occur simultaneously: first it sends an action potential out along its main axon to some distant cortical or subcortical target; at the same time, the axon collaterals of the cell generate a local "dynamic pattern" which radiates out across dozens of surrounding cortical columns.
Each of these adjacent columns also contain pyramidal cells whose apical dendrites pick up the collateral activation from the firing pyramidal cell. This allows the firing cell to influence the activity level of the receptor cells, presumably increasing the probability of those cells firing, since the axon collaterals are excitatory. The overlapping patterns generated by local networks of thousands of synaptic junctions would serve to couple the activities of the participating neurons (viz. Kisvarday et al. [49]), generating field interactions that could spread across the cortical neuropil.
These field interactions, or dynamic patterns, are also capable of directly influencing the activity levels of individual pyramidal cells within cortical columns. Thus, the interaction of the columnar and tangential systems offers a plausible mechanism for the simultaneous processing of local and global information within the cortex. Moreover, this architecture suggests a basis for globally integrated "motor representations" guiding volitional acts. Pyramidal neurons in motor cortex, as much as neurons processing perceptual inputs, are part of this tangential network.
We would emphasize that the tangential network appears to consist not only of the continuous mesh of pyramidal axon collaterals and apical dendrites, but a massive system of cortico-cortical afferents which interconnect nearly all areas of the cortex [21] (see Fig. 2). This cortico-cortical system includes fibers making up the corpus callosum. In concluding a recent review of the literature on "aspects of cortical connectivity", Jones [51] writes,
One interpretation of this terminal pattern would be that the cortico-cortical fibers are forming terminal sites primarily on the apical dendrites and basal dendrites of layer III pyramidal cells, as these dendrites ascend through layers I-III and descend into layer IV, respectively. [p. 162]
The excitatory nature of the tangential network would suggest that neural excitation from a local focus could spread, unchecked, across the cortex. In actuality, inhibitory influences prevent this from occurring, at least in healthy brains (the tangential fibers in layers III-IV (Fig. 2) could play a role in inhibiting the spread of excitation in the network). In epileptics, however, inhibitory mechanisms are overridden and a "storm" of neural activation, or seizure, occurs. Patients with severe seizure disorders have been shown to benefit greatly from an operation in which the corpus callosum is severed. As Sperry [52] commented concerning a group of such "split-brain" patients, the operation was,
surprisingly successful; the seizures were not only prevented from spreading, but were almost eliminated from both hemispheres. Moreover, upon recovery there was no noticeable alteration in the patient's overall intelligence, personality and character. [p. 16]
Despite these remarkable findings, Sperry [52] concluded that severing of the corpus callosum resulted in,
the functional disengagement of the right and left hemispheres with respect to nearly all cognitive and other psychic activities .... Each [hemisphere] ... seems to have its own conscious sphere [of] ... mental activities and the whole realm of gnostic experiences of the one is cut off from the corresponding experiences of the other hemisphere with only a few exceptions. [p. 17]
Another feature of this tangential cortical network that may suit it for the global representation of neural activity is its likely role in the spread of activation initiated by subcortical systems involved in arousal and attention. These activation systems are known to project predominately to the superficial (and deepest) layers of the cortex, in contrast to the specific thalamocortical pathways, which project predominately to the middle layers [36, 37, 53, 54].
A review of "the organization and evolution of nonspecific thalamocortical projections" by Herkenham [37] is interesting in this regard. He notes that for many years neuroanatomists have assumed that the source of thalamic projection fibers terminating in the superficial layers of the cortex was the traditionally defined nonspecific thalamus. Recent non-primate studies, using HRP-tracer techniques, have revealed a more complex cortical distribution. The targets of the nonspecific (intralaminar) nuclei have been found to be predominantly neurons in the deepest layers of the cortex. Superficial, or "Layer I", projections, according to Herkenham, can originate from any thalamic nucleus, but have their richest sources in "paralaminar" regions adjacent to the intralaminar complex (ILC). His research suggests a medial to lateral continuum from nonspecific (deep layer/intralaminar) through mixed nonspecific/specific (many Layer I/paralaminar) to specific (mainly mid-layers/sensory relay) nuclei. Herkenham's [37] "Tripartite Division" of the thalamus could provide a taxonomy taking research beyond the rigid distinctions between specific and nonspecific thalamic nuclei which, as Goldman-Rakic [7] has argued "may no longer be appropriate."
Paralaminar projections to the superficial layers of the cortex could globally influence the spread of field interactions in the tangential network in a "mixed" (local/global) but multimodal fashion. Deep layer, intralaminar projections could have even more global influences upon the network, possibly via the "long surface parallel terminal axons" which Szentagothai [21] describes. The sources of these layer I terminal axons appear to be neurons whose cell bodies are located in the deepest layer
of the cortex (see Fig. 2) This could resolve the seeming dilemma posed by Macchi and Bentivoglio's [44] findings (cited above) that intralaminar projections have restricted fields and collateral spread. The intrinsic architecture of the tangential network might serve to spread nonspecific activation.
Herkenham [37] cites several physiological studies that support this scenario. They found that stimulation of ILC nuclei induces rhythmic oscillations "simultaneously in superficial and deep layers, or often earlier in deep layers." [p. 424]. Recent experiments with cats also found oscillatory neurons to be con-centrated in the upper and lower layers of the cortex. Gray et al. [55] report that few of the neurons which show stimulus dependent oscillations are found in layer IV. These newer findings are particularly pertinent because the experiments were designed to study the possible mechanisms of visual attention.
The microstructure of this tangential network is only imperfectly understood at present, and further research will undoubtedly alter some of the anatomical conceptions offered above. Recent research by von der Malsburg and Singer [50], however, supports the concept of field interactions generated within the neuropil of the neocortex serving as a basis for the global integration of neurocognitive activities. This research is reviewed, at length, in the next section.
5.2 The Functions of the Tangential Intracortical Network (NGW).
We have argued that most of the cognitive processing carried out by the central nervous system can be understood in terms of the automatic and relatively autonomous activities of large collections of specialized processors. But such parallel, distributed systems have inherent limitations. They are not designed to adapt to novel situations which require the inhibition of automatic routines, and the creation of new functional connections between different specialized processors. Moreover, they are unsuited for more conscious kinds of problem-solving, where attention must be selectively and flexibly focused in accordance with intentions and plans. Finally, it is difficult to imagine how they could generate the globally integrated representations of ongoing experience (and the memories of such experience) that so strongly characterize our conscious awareness.
Global workspace theory hypothesizes that to accomplish these conscious functions, the nervous system requires a global "representational space". This global workspace must be readily accessible to all of the specialized processors, and allow for their mutual interaction. Coalitions of processors must be able to compete for access to the workspace, as well as cooperate with other coalitions such that a single coalition of processors can dominate global computations at that moment, creating a unified state of awareness. This GW allows for the integrated, central representation of multifarious cognitive activities. Its architecture addresses the binding problem of how unified conscious percepts can emerge out of the activities of multiple specialized processors [4, 8, 15].
The NGW model proposes that an explicitly definable neural network subserves the functions of such a global workspace. This network corresponds to the "tangential system" [21] described above. In keeping with more recent terminology, we will refer to this system as a "Tangential Intracortical Network", or "TIN". Recent work by von der Malsburg and Singer [50] on self-organization and pattern formation in cortical networks provides a considerably more detailed appreciation of the nature of this network.
In conceptualizing the functions of this tangential network, Von der Malsburg and Singer [50] have looked to examples of field interactions studied in physics and chemistry, where extremely large numbers of atoms or molecules interact with each other. They note that the cerebral cortex contains on the order of 1014 synapses. In physical systems of densely packed units, where the activity of adjacent units is mutually influencing, several principles of self-organization operate. These can be summarized as follows:
1) Local interactions tend to self-amplify. Synchronous or correlated interactions among coupled units strengthen and spread across ensembles of units, creating coherent activity patterns.
2) Developing activity patterns compete. The strongest (most coherent) patterns vigorously grow at the expense of others. This leads to the formation of activity "domains" of different self-amplifying patterns.
3) Domains of activity tend to cooperate. In spite of the overall competition in the system, domains of correlated activity will tend coalesce to form larger, coherent activity patterns. If there are no outside influences acting on the system, the activity patterns with the most internal cooperativity and the least competition will win out.
(after von der Malsburg and Singer, 1988)
Von der Malsburg and Singer [50] note that a fundamental correlate of these three basic principles is that "global order can arise from local interactions .... ultimately leading to coherent behavior." [p. 71] In a large self-organizing network, a number of competing local domains can co-exist, but the tendency of the network is generally towards attaining a globally ordered state. Because of this, even a relatively weak stimulus towards global organization can decisively influence developing local patterns. Let us look at how these "Principles of Cortical Network Organization" can be applied to mechanisms of "self-organizing interactions in the striate cortex."
It is common knowledge that the primary visual pathways of the nervous system project, in an highly ordered topography, upon striate cortex. Contemporary research on vision [56] has shown that visual signals from the retina are analyzed by highly specific "feature detectors" that form alternating, modular domains (resembling zebra stripes in experimental preparations of cortex). These domains are each roughly topographic, i.e. the spatial relationships between points in the original visual scene are maintained. But, of course, such topographic maps bear little resemblance to what we consciously perceive as a unified gestalt of shapes, colors, motion, etc. All of these features of the visual scene are processed separately by specialized areas in visual cortex. Thus, even a simple scene generates dozens of topographic maps in the various areas of visual cortex.
The questions that von der Malsburg and Singer [50] have been exploring are how do these topographic maps arise, and how are their multiple representations integrated in visual perception? They begin their exploration with a discussion of how topographic maps are established during embryonic development and early life, in other words, how the initial "hardwiring" of the visual system proceeds from the retina to subcortical nuclei and ultimately to the visual cortex. These developmental processes, as would be expected, produce axonal connectivity patterns that tend to be parallel and modular.
Even the local, tangential intracortical networks of axon collaterals and apical dendrites self-organize their cells into "neuron clusters" that tend to be spaced periodically, contributing to the formation of local "feature domains". These local domains, since they are topographically organized and "tuned" to detect particular coherencies in the visual scene, spontaneously produce multiple, quasi-discrete topographic maps in response to visual inputs (Principle 2). But self-organization does not stop at this local, topographic level. Tangential connections "spanning between neuron clusters whose activation patterns show some statistical correlation are stabilized by coherency matching", creating more global maps representing two or more feature domains [Ibid, p. 91].
Thus, by the simple reiteration of the very same processes of self-organization which ... lead to the establishment of maps encoding for topographic neighborhood relations, it is possible to generate nontopographic maps which represent dimensional neighborhood relations in feature space. [Ibid, p. 92]
In this manner, increasingly global representations of visual information can be generated. These representations are "more global" not only because they integrate feature domains, but because they continually compare real-time representations with past representations: "Those which best match the already established connectivity pattern will have a competitive advantage". [Ibid, p. 92]
What makes this tangential network particularly adaptive is that multiple topographic and nontopographic maps can co-exist within it at the same time. "Thus, one of the basic principles of cortical organization appears to be allowing for the parallel computation of both local and global correlations within the same representation" [Ibid, p. 93]. These dynamics suggest a network in which specialized processors are continuously competing and cooperating to form increasingly global representations, at the same time that local feature detectors are carrying out their preliminary analysis of incoming visual information. It should be kept in mind that the "global correlations" the authors refer to are confined to striate cortex and, thus, represent integrations of very basic elements of vision. The tangential intracortical network, however, is not confined to striate cortex, but spans the entire cortex. And, as von der Malsburg and Singer [50] conclude,
Since the basic principles of cortical processing can be expected to be the same in all cortical areas, it will become a fascinating challenge to both theoreticians and experimentalists to examine the functional properties that emerge if these processes of self-organization are iterated beyond primary sensory areas and applied to architectonics in which "dimensional" or "conceptual" vicinity is ever more emphasized at the expense of topographical vicinity. [p. 95]
It is significant that even at the level of feature detection accomplished within striate cortex, interactions between clusters of feature-specific neurons play an important role in pre-attentive processes, providing "the basis for any pre-attentive segmentation of scenes and 'figures' and the necessary prerequisite for any subsequent identification of patterns." [Ibid, p. 92]. Again, the nature of the tangential network is such that these low-level attentional mechanisms could be reiteratively invoked at higher levels of information processing.
Finally, von der Malsburg and Singer [50] point out that the central mechanism in all of these activities is most likely the selective weakening and strengthening of the myriad synaptic connections of the tangential network as incoming neural information and "central core systems" continually modify the state of cortex. Thus, their work has direct relevance to connectionist modeling of perception and learning.
Most of the prerequisites for learning mechanisms would be fulfilled: The modification of the coupling strength of the neuronal connections are activity dependent; the modification rules are based upon local contingency matching and hence have the ability to associate contiguous events in time and space; the occurrence of modifications is gated by central core systems and hence can, in principle, be made dependent upon global states such as arousal, attention and motivation; and finally, modifications are long-lasting and can thus serve to establish permanent representations.... [Ibid, p. 94]
This paper pursues the hypothesis that the self-organizing principles for cortical networks elucidated by Von der Malsburg and Singer are essentially those required for a global workspace. Thus, the tangential intracortical network they describe could provide an essential basis for the representation and processing of conscious information. Having established an anatomical basis for a global workspace, we can move on to describing the influences of cortical attentional mechanisms and "central (or reticular) core systems" upon this NGW.
6.0 Nucleus Reticularis and the Control of Global Attentional Processes.
The neural substrates of attention, like those of consciousness, consist of widely distributed systems operating at both cortical and subcortical levels of the brain [11, 26-30, 57]. The complexity of these systems, like that of the intracortical network described above, is daunting. And while any model which attempts to capture the basic elements of such systems represents an over- simplification, the very complexity we are dealing with makes simplifying models essential.
The sections to follow will offer such a neural model for global attention. The model does not attempt to encompass all aspects of the problem (e.g. we generally avoid discussions of preattentive, unimodal processes, and the role of motivational and emotional states). Rather, it deliberately focuses upon global perceptual and cognitive processes. The central structure in this neural model is the nucleus reticularis (NR) as the heart of an "extended reticular-thalamic activating system" regulating the flow of information between the cortex and thalamus. Other subcortical structures clearly play important roles in attentional processes, particularly other thalamic nuclei [11, 25, 30], but none has been implicated in as many neural processes related to the regulation of information flow and rhythmic EEG activity in the CNS [37, 40, 54, 58].
To review the earlier discussion (p. 8-9), the reticular nucleus consists of a thin sheet of specialized neurons which covers the surface of the thalamus both laterally and anteriorly. NR is thinnest in its lateral-posterior extension. The rostral, or anterior, pole of NR is considerably thicker. NR does not send any projections to the cortex, but nearly all of the reciprocal path-ways running between the cortex and principal nuclei of thalamus pass through NR, and a significant proportion of their axons give off collaterals as they pass through NR. Nonspecific projections from the intralaminar complex (ILC) are also known to give off extensive collaterals. An exception to this pattern is a tract of fibers originating in prefrontal cortex which projects directly to the anterior pole of NR, and to MD (refer to Fig. 1) [36, 37, 40].
The reticular nucleus has extensive reciprocal connections with ILC, as well as the midbrain reticular formation (MRF), making it an integral part of the reticular core of the brain [36, 37, 40]. The bulk of NR's projections, however, are to the ventral (principal) nuclei of the thalamus which lie directly interior to it (refer to Fig. 1) Thus, one of its primary functions would appear to be the modulation of these specific thalamic nuclei [39, 40, 58].
In Figure 1 the gross anatomy of the thalamus and the midbrain are portrayed, schematically. The two hemispheres have been "removed" to reveal a cross-sectional view of these subcortical structures. While the figure is designed to give a sense of the anatomy of these structures, it has been necessary to exaggerate certain features to illustrate important points. Obviously, the hemispheres are shown greatly reduced in size relative to the subcortical structures. On the other hand, the thin layer of cells forming the reticular nucleus ("NR", left and right) have taken on an exaggerated thickness in the figure (relative to the other nuclei of the thalamus) in order to illustrate connectivity patterns with the midbrain, thalamus and cortex.
It is common knowledge that virtually all information conveyed to the cortex from visual, auditory and somatosensory pathways is relayed through the thalamus. The thalamus is also the predominant source of internally generated activation of the cortex, both specific and nonspecific. As noted above, nearly all thalamocortical pathways pass through nucleus reticularis on their way to the cortex. A significant proportion give off collaterals. The same is true for reciprocal pathways running from the cortex to the thalamus. This architecture places NR in an ideal location for the central control, and representation, of information flow between the thalamus and cortex.
Studies have shown that neural volleys ascending along thalamocortical axons can generate activity within NR that immediately suppresses the flow of impulses along the ascending axons [59 - 62]. These and other findings have led neurophysiologists to characterize the reticular nucleus as consisting of a vast array of tiny gates, or "gatelets" [58], controlling the flow of information to the cortex [40, 57, 58]. These intricate gating mechanisms most likely serve a variety of purposes. In a subsequent section we will discuss evidence for their role in the synchronization of EEG activity.
6.1 Three Global Attentional Systems.
The NGW model hypothesizes that the activation of NR neurons by the axon collaterals of these thalamocortical and corticothalamic fibers plays an important role in global attentional pro-cesses. By this we mean that NR is central to the "control of access" of specialized processors to the cortical GW (p. 6). In particular, we posit that the corticothalamic pathways which pass through NR as they converge upon the ventral thalamus, modulate their own information processing activities via these NR "gatelets". These corticothalamic pathways originate, for the most part, in the posterior cortex, where sensory information is initially cognized.
In the present model, this "Posterior Cortical" (PC) system constitutes one of three major attentional sub-systems converging upon NR. It corresponds, roughly, to models of posterior cortical attentional systems proposed by Mesulam [30] and Posner [28]. However, PC constitutes a more distributed attentional system than suggested by these earlier models, since it consists of cortical fields (and their projections to the thalamus) from the parietal, occipital (and portions of) the temporal cortex. The key anatomical feature of PC is the collaterals given off by its cortical efferents as they pass through NR. To briefly characterize this attentional system in functional terms: it selectively integrates sensory inputs based upon relevant past experience and present contingencies. We might call this the function of "immediate attention", in contrast to attention guided by such things as goals and intentions. We will elaborate further upon these distinctions in the section to follow.
A second major attentional subsystem is found in the prefrontal regions of the cortex (the cross-hatched area in upper half of Fig. 1). It sends the bulk of its thalamic projections to the medial dorsal nucleus (MD) (cross-hatched area, lower half of figure), but sends a major tract to the rostral pole of the reti-cular nucleus as well. Prefrontal cortex projections to NR are direct, not collateral, and have pervasive effects upon activity throughout the nucleus. This prefrontal-mediothalamic system is capable of activating large portions of NR. We will refer to this system simply as "PfC". Animals studies involving stimulation of the prefrontal-mediothalmic tract have demonstrated that the effect of stimulation is to "close down" NR gatelets, inhibiting the flow of sensory information to the posterior cortex [40, 58, 62].
While generalized PfC activation has the effect of blocking all modalities of sensory input to the cortex (by generating synchronous EEG activity in the 7-13-Hz range), Skinner and Yingling [40, 62], in experiments with cats, were able to demonstrate selectivity in the effects of PfC upon NR. They did this by cryogenically blocking different portions of the prefrontal-mediothalamic tract and observing the effects upon cortical event-related potentials in the primary visual and auditory areas. What they discovered was that, depending upon which portion of the tract was blocked, a "gate" was opened in the reticular nucleus allowing sensory input to reach the cortex in a single modality. Thus, PfC would appear to be able to differentially close NR gatelets in response to selective patterns of activation within the prefrontal areas.
These findings have obvious relevance to the understanding of the neural mechanisms of attention. Prefrontal cortex is known to mediate a number of functions related to higher forms of cognitive activity [63, 64], and both primate and human research have suggested a role for PfC in more purposeful forms of attention [29, 30, 65]. As discusssed in the next section, Goldman-Rakic [65] has hypothesized a central role for prefrontal cortex in the processes of working memory. And in agreement with her theory, the present model considers PfC as essential to a "secondary", or reflective, consciousness able to selectively impose its plans and purposes upon the more stimulus-bound forms of information processing carried out in posterior cortex. Again, elaboration upon these systems/functions will be offered below. But a complete appreciation of NR's central place in the global attentional matrix requires the inclusion of a third, converging system critical to "alerting" [29] and the orienting response [38, 40, 57, 58].
The effects of both PC and PfC projections upon NR are highly selective. These systems would appear to be able to inhibit the flow of information to the cortex in very specific ways. These more localized functions as commonly referred to as "phasic" activation, in contrast with more generalized excitatory influences upon cortical activity which are termed "tonic". While phasic activation, via PC and PfC, can serve to filter out sensory information that is not relevant to the focus of attention, it does not have the capacity (at least directly) to enhance the ascending flow of neural impulses coming from the sensory systems and thalamus [40, 41, 57, 58].
This capacity resides in the ascending reticular activating system, which originates in the brainstem reticular formation [38]. The reticular formation receives inputs from wide areas of the cortex. In a similar fashion to NR, it receives extensive collaterals from the major sensory tracts as they pass near or through it on their way to the thalamic relay nuclei (refer to Fig. 1, p. 8). More importantly, activation of the reticular formation inhibits neurons in nucleus reticularis (which are themselves inhibitory), "opening up" NR gatelets [40, 58]. This is the likely mechanism by which increased activity in the retic-ular formation leads to the "facilitation of trans-thalamic sensory transmission" [30, p. 135].
Projections originating in the midbrain reticular formation (MRF) form the third converging influence upon NR gatelets. Ana-tomical studies reviewed by Scheibel [58] has shown that these projections arise predominately from two sets of nuclei: the superior colliculi ("sc") and the cuneiform nucleus ("cun"). These reticular core nuclei appear to contain a representational map of "the three-dimensional spatial envelope surrounding the organism" [58, p. 62]. This "spatial envelope" is very likely central to the orienting response mediated by MRF [38, 57, 58].
It is commonly thought that the tonic activation initiated by MRF has only very diffuse effects upon information processing. This conception is belied, however, by more recent physiological and anatomical findings. Bloom [66], in a recent review of role of the brainstem activating systems in cortical function, concludes that their effects are not just tonic and generalized, but "the global afferent produces conditions that both enhances and differentiates the specific local responses .... [thus providing] a means for the selective regulation of responsivity. [sic]" [p. 416]. Steriade and Llinas [54] review recent physiological findings on the thalamic extension of the reticular core supporting Bloom's conclusions. A more differentiated appreciation of the anatomy of these core structures also supports this view. Of particular relevance to the present model is research reviewed by Scheibel [58] that projections from the MRF,
sweep forward on a broad front, investing midline thalamus [ILC] and nucleus reticularis thalami [NR]. The investiture is precise in the sense that the [sc/cun] sites representing specific zones of the spatial envelope (receptive field) project to portions of the nucleus reticularis concerned with similar peripheral fields via projections from both sensory thalamus and sensory association cortices. [p. 62]
Thus MRF, like the two cortically-based subsystems of the present model, appears to have the capacity to differentially modulate the activity of NR gatelets. Summarizing several decades of research on this attentional matrix, Scheibel [58] builds upon Skinner and Yingling's [40] PfC-NR model, integrating "sensorimotor cortex" (PC) into the attentional matrix. He concludes,
From these data, the concept emerges of a reticularis complex [NR] selectively gating interaction between specific thalamic nuclei and the cerebral cortex under the opposed but complementary control of the brain stem reticular core [MRF] and the frontal granular cortex [PfC]. In addition, the gate is highly selective; thus, depending on the nature of the alerting stimulus or locus of central excitation, only that portion of the nucleus reticularis will open which controls the appropriate subadjacent thalamic sensory field. The reticularis gate [thus] becomes a mosaic of gatelets, each tied to some specific receptive zone or species of input. Each is under the delicate yet opposed control of (a) the specifically signatured sensory input and its integrated feedback from sensorimotor cortex [PC], (b) the mesencephalic reticular core [MRF] with its concern more for novelty (danger?) than for specific details of experience and (c) the frontal granular cortex-medial thalamic system [PfC] more attuned to upper level strategies of the organism, ... Perhaps here resides the structurofunctional substrate for selective awareness and in the delicacy and complexity of its connections, our source of knowing, and of knowing that we know. (p. 63)
As the heart of this attentional matrix, the nucleus reticularis mediates the influences of PC, PfC and MRF upon attention, as well as modulating the activities of the principal nuclei of the thalamus.
6.2 An Extended Reticular-Thalamic Activating System.
In the introduction we noted that psychologists have traditionally portrayed consciousness as an essentially serial (although highly integrated) process, involving limited information processing capacity and high error rates. Yet, more recent cognitive and neuroscience research has suggested that most information processing is carried out by highly specialized, modular processors. In this paper we have pursued the hypothesis that this apparent contradiction in perspectives reflects a certain large-scale architecture of the nervous system, one that is functionally equivalent to a "global workspace in a system of distributed specialized processors" (p. 3).
According to the NGW model, at the core of this unifying, large-scale architecture is the nucleus reticularis of the thala-mus, allowing the cortex to "gate" it own inputs, thereby giving the nervous system the ability to limit its information processing load. In this context, the limited information processing capacity associated with conscious cognition becomes a virtue of the system. Besides its gating functions, the nucleus reticularis is an integral part of an extended reticular-thalamic activation system regulating global cortical EEG activation.
The model hypothesizes that the tangential intracortical network (TIN) is the locus of representation for this limited, and highly coherent, stream of information processing. Its activities serve both: to create increasingly global representations of underlying modular processes; and to spread rhythmic activation patterns, modulated by the reticular-thalamic core, across the cortex. As von der Malsburg and Singer [50] concluded in describing the physiology of TIN: "the occurrence of modifications" in the network depends not only upon "local contingency matching", but "is gated by central core systems and hence can, in principle, be made dependent upon global states such as arousal, attention and motivation." (p. 20)
These central core systems are essentially those described by Moruzzi and Magoun [38], extending from the brainstem reticular formation, through the nonspecific thalamus, to the widely distributed projections of the intralaminar complex upon the cortex (and likely including paralaminar "nonspecific" projections as well, cf. discussion p. 15) . What has been too frequently ignored by critics of the concept of a unified activating system is the extended nature of this system, including descending, cortical influences upon it. To quote Magoun [67],
In its ascending and descending relations with the cerebral cortex, the reticular system is intimately bound up with and contributes to most areas of nervous activity. ... and through its relations ... with the cerebral cortex, with most of the central integrative processes of the brain. (p. 10)
Clearly the cerebral cortex has a central role in the medi-ation of attentional processes. Two of the three converging, attentional sub-systems we have described in previous sections are centered upon the cortex. But we argue that the mechanisms by which these cortical structures are able to influence their own attentional states only become clear when the influence of thalamic projections upon the cortex are taken into account. It is these ascending systems, according to the present model, that globally activate the tangential intracortical network.
Included in these ascending systems is the midbrain reticular formation, and its highly differentiated projections upon nucleus reticularis (previous section). Taken together, the "opening" and "closing" influences of MRF and PfC upon NR gatelets offer a viable basis for understanding how the brain selectively regulates the flow of information to the posterior cortex (PC), allowing conscious awareness to vary along a broad continuum from generalized alerting to focused attention. But PC is not merely the passive recipient of this selective flow of information. An extensive body of evidence from both clinical studies and experiments with monkeys have demonstrated the importance of the posterior cortex in certain forms of attention [30, 28, 68]. Luria [63] characterizes the human posterior cortex as a "unit for receiving, analyzing and storing information" [p. 67]. It takes inputs from the sensory relay nuclei, synthesizes them into mental representations of ongoing experience, and integrates these representations with previously learned information. Luria contrasts these more reactive, stimulus-bound forms of information processing with those mediated by the prefrontal cortex which, "creates intentions, forms plans and programmes of ... actions, and regulates ... behavior so it conforms to these plans and programs, ...." [p. 79-80]
While information processing in the posterior cortex is more automatic and, thus, less intentional than processing mediated by PfC, it clearly entails the capacity to selectively attenuate/amplify portions of the attentional matrix. This is not surprising given the extensive reciprocal projections from nearly the entire posterior cortex to NR and the principal nuclei of the thalamus. In characterizing PC's effects upon the conscious processing of information, it is important to note that the "association areas" of the cortex appear critical to more conscious forms of attention. Experiments in the neurophysiology of attention indicate that sensory inputs which reach the primary cortical areas do not necessarily become conscious. Neurons in primary cortex automatically respond to any liminal sensory stimulus in their modality, and will continue to do so as long as the stimulus is presented. This remains the case whether the stimulus is attended to or not [57, 69]. Knight [70] reviews human studies of neural activation associated with conscious attending indicating that,
the first recordable difference between attended and non-attended stimuli occurs at 60 to 80 msec post-stimulus ... long after the peripheral sensory volley has reached primary cortical receiving areas." [p. 333]
Work by Libet [71, 72], involving surgical patients with implanted electrodes, suggest that it can take as long as a half-a-second for stimulation of primary cortex to become conscious. These findings imply that a complex, multiply-determined series of neural events precede the registering in consciousness of even simple sensory information. Such a scenario is consistent with the concept in GW theory of the global broadcasting of a conscious, "global message" being preceded by an input stage in which large numbers of specialized processors compete (and co-operate) to gain access to the global workspace.
We hypothesize that the extensive system of reciprocal projections running between the thalamus and cortex -- under the control of NR gating mechanisms -- are central to carrying out such a series of hierarchically organized processes of selection necessary to the input stage of conscious perception. This selection process can be thought of as a complex filtering of inputs (over a 60 to 500 msec period) from a diverse array of specialized processors. Cooperating processors gradually accrue greater levels of activation, increasingly inhibiting the activity of competing processors, until their integrated output predominates in the neural global workspace. Once this condition is reached, the output of the "victorious" coalition of processors is broadcast over wide areas of the cortex. As discussed in earlier sections, the TIN plays an integral role in these processes.
6.3 Prefrontal Cortex as an Executive Attentional System.
How might we further characterize the functional role of prefrontal cortex in this model? We noted in the previous section that connectivity patterns from prefrontal cortex to NR are very different from those between posterior cortex and NR, as are the contributions this frontal-thalamic system makes to global attention. In cognitive terms, PfC acts as an "executive system", controlling activities related to the accessing of schemas, plans and voluntary goals. Neisser [73] defines several properties characteristic of such a system: attention is serial and focused; perception is active, combining ongoing percepts with the recall of previous experience and abstract knowledge; and the entire process is goal-directed. He contrasts such "secondary process" activities with "primary process" which is "a multiple activity, somewhat analogous to parallel processing in computers ...." [p. 304] Finally, a defining property of an executive system is that it acts upon other sub-systems using their inputs for its particular purposes. Compare this with Fuster's [64] final summary in his book The Prefrontal Cortex,
The prefrontal cortex is thought to be essential for the synthesis of cognitive and motor acts into purposive sequences. This synthetic function is presumably accomplished with the support of three subordinate functions: (1) an anticipatory function ensuring the preparatory set of sensory and motor systems; (2) a retrospective function of provisional [i.e. task-related] memory; and (3) the suppression of external and internal influences ... that interfere with the formation of behavioral structures. [p. 145]
At the most basic level, the prefrontal cortex selectively activates NR gatelets, blocking the flow of irrelevant stimuli to the posterior cortex (3). But this only describes the final outcome of a series of cognitive processes which occur in prefrontal cortex prior to selective gating. As Mesulam [30] points out, this "attentional filtering occurs at a quite advanced stage of information processing, since the unimodal sensory areas whose inputs are blocked out of conscious awareness continue to be active." [p. 139]. All of this suggests that the prefrontal cortex makes judgments of stimulus "irrelevance" based upon complex, anticipatory cognitive sets (1).
Thus the capacity of the prefrontal cortex to focus attention upon highly relevant stimuli, is guided not so much by immediate perceptions and needs (more a function of PC and MRF), as by purposes and plans [63, 64]. PfC not only allows attention to be focused in the moment, but over time (such as when we are searching for a particular stimulus in a complex array of stimuli, or anticipating the occurrence of a stimulus). Vital to this focusing process is the ability of prefrontal cortex, through its projections upon NR, to inhibit the orienting response [40, 63]. In a recent paper on "The Prefrontal Contribution to Working Memory and Conscious Experience", Goldman-Rakic [65] presents evidence from primate research supporting Fuster's [63] conclusion that a special form of memory is associated with the functions of PfC. Her research suggests that PfC neurons show a,
data-holding capacity such as has been ascribed to working memory [which] would allow information coming in at one point in time to be temporarily associated with information occurring after many seconds. [p. 16]
Goldman-Rakic uses the term "representational knowledge" to describe information held in working memory. She contrasts the capacity of prefrontal cortex to generate temporary, internal representations - independent of ongoing perceptual processing - with conditioned learning, which simply associates immediate sensory experience to pre-existing response and reward contingencies. Indeed, the evolution of a capacity for representational knowledge may provide a mechanism for overriding the governance of behavior by conditioning or by reflexive or innate or prepotent responses whenever such behavior could be maladaptive for the organism. [Ibid.] She suggests that the role of PfC in the internal generation of provisional memories may extend not just to representations of recent experience, but include "the ability to bring past experiences to mind for the purposes of contemplation, thought or action ...." [p. 18]. Goldman-Rakic [65] notes that while the mechanisms for holding and recalling experiences and knowledge in working memory are presently unknown, it is very likely that they involve "a cooperative relationship between the hippocampus and the prefrontal cortex." [p. 12]. A we have noted previously, extensive connections exist between the prefrontal cortex and medial dorsal nucleus (MD) [64]. Studies have consistently shown that damage to the medial thalamus, and particularly MD, result in deficits in memory. Indeed, it is very likely that MD serves as a crucial "bridge" between PfC, the hippocampus, and other structures central to memory processes [74].
As for NR, the evidence presented previously in this paper strongly implicates this thalamic nucleus in PfC's capacity to inhibit the orienting response, and "gate" information processing activities in PC. It is our contention that prefrontal cortex acts as an executive attentional system by actively influencing information processing in the posterior cortex through its effects upon the nucleus reticularis. In this manner, the highly parallel, "primary process" functions of the posterior cortex are brought into accord with increasingly complex and intentional cognitive schemes generated within the prefrontal regions of the brain.
7.0 The Binding Problem Revisited - The Integrated
Representation of Conscious Perceptions.
The focus of the NGW model in the past three sections has been upon the attentional matrix formed by the cortical and midbrain systems converging upon nucleus reticularis thalami. Ironically, it could be argued that this tri-leveled architecture (cortex, thalamus, brainstem) is the quintessence of a parallel-distributed system. The limitation inherent to parallel, distributed neural systems is that widely separated processing units are poorly suited to the integrated representation of perceptual experience which is a defining characteristic of conscious awareness. We referred to this dilemma in the introduction as the "binding problem", that is: how are the activities of all these widely distributed areas coordinated in the service of selective attention and conscious perception?
As discussed earlier, the traditional approach to this problem has been to try and discoverer the area(s) of the CNS where the most extensive convergence of polymodal inputs occurs. But as knowledge of each candidate area (e.g. association cortex, nonspecific thalamus, etc.) has become more differentiated, highly parallel, modular micro-architectures have been increasingly revealed. In recent years, it has been suggested that the focus needs to shift from input channels to "reentrant", or feedback, connections which are found throughout the cortex and between it and subcortical nuclei, particularly the thalamus [11, 12, 56]. The general thinking it that these reentrant pathways could: 1) link together the functions of interacting areas and/or; 2) funnel highly processed and integrated information from "association" cortex back to primary sensory areas, producing highly integrated representations.
A third alternative, presented in this paper, is that the tangential intracortical network, in conjunction with central core systems, accomplishes these functions. All three scenarios find some support in the research literature and each is likely to play some role in neural integration. In Baars and Newman [15] the combining of these scenarios is explored, within the framework of Global Workspace theory, in terms of the representation of visual percepts in consciousness. Here, we wish to explore the alternative hypothesis that binding is ultimately a function of the rhythmic correlation of global cortical activation via an extended reticular-thalamic activation system.
* * * * *
Conscious percepts are especially likely to consist of novel combinations of perceptual features. Moreover, they arise rapid- ly and are largely transient in nature. A viable theory of conscious perception must suggest some material basis for understanding how such transient conjunctions could arise in the nervous system. William James [23] coined the metaphor "stream of consciousness" to capture the evanescent and ever changing nature of this process, and rightly concluded that consciousness cannot be understood in terms of fixed structures, but as a highly transient process. The integration of features constituting a conscious percept, as Crick [34] first hypothesized, must give rise to some sort of transient binding of widely distributed networks of neurons [13, 14, 32].
A possible example of such transient binding is the "coherency matching" described by von der Malsburg and Singer [50], in which the activities of networks of neurons are linked by momentary correlated oscillations (see pp. 18-20). As noted already, recent animal studies [33, 55] support this idea, showing that neurons in the primary visual cortex activated by the same object in the world tend to discharge rhythmically and in unison. Coherency matching would appear to be a natural, self-organizing property of densely interconnected neuropil, generating field interactions that lead to the spontaneous integration and representation of increasingly global neural states. The present model hypothesizes that the tangential intracortical network (TIN) described in previous sections possesses such "self-organizing properties" by which "global order can arise from local interactions." (p. 17).
But a more globally distributed, integrative impetus is re-quired if locally generated coherencies, originating in widely separated areas of the cortex (e.g visual, auditory and somatosensory areas) are to be transiently combined to form unified percepts. Moreover, the idea that the binding of cortical processors could, by itself, generate conscious percepts is at odds with what is known about the nature of consciousness and attention. Only a system capable of the coordinated binding (and unbinding) of widely separated modular units - guided by alternating considerations of novelty, stimulus relevance, volitional intention, etc. - could satisfy all the conditions for conscious perception. This is why, in our view, we must look outside the cortex itself for a solution to the binding problem.
The key to the solution of the binding problem lies in an understanding of the functions of the nucleus reticularis as the heart of an extended cortical activation system. Only when the activities of the TIN are subjected to the global influences of the reticular core do the "local contingency matchings" of this cortical network coalesce into the unified representations that characterize conscious perception. The role NR plays in generating such unified representations can be heuristically divided into two general functions: 1) it controls access to the NGW (previous section); and 2) it integrates relevant inputs into "global messages" (see introductory sections) suitable for "broadcast" by this activation system.
Thatcher and John [57] summarize a variety of EEG studies, employing multiple electrodes, which have shown that EEG rhythms are "in constant motion shifting across the cortical surface in complex recurring patterns." These scanning phenomenon have been studied in the greatest detail for alpha rhythms (7-13 Hz). In general, alpha waves are most prominent in the visual cortex, but are found in varying distributions throughout the cortex. Alpha is commonly thought to be an index of resting activity in an awake subject, when 70% of the EEG is at alpha frequencies. But what these EEG scanning studies indicate is that alpha activity sweeps periodically across large domains of the cortex in moving waves. The speed of scanning has been shown to vary as a function of general levels of EEG arousal, increasing as arousal levels increase. And although alpha activity, in itself, appears to be an index of a lack of cognitive processing, such period fluctuations in alpha may reflect filtering effects of the attentional matrix described above. Thatcher and John wrote,
The concept of scanning was originally introduced by Pitts and McCullough (1947) in relation to perception in humans. Their idea of a periodic subcortical scanning pulse that probes the momentary state of neural excitability was consistent with evoked potential recordings demonstrating excitability fluctuations coincident with background EEG. For instance, Bartley in 1942 demonstrated a periodic alteration in the excitability cycle of the sensory evoked potential coincident with the frequency and phase of the alpha rhythm. Similarly, Dustman and Beck (1965) and Morrell (1966) demonstrated that behavioral reaction time was minimal when the triggering stimuli occur on the rising phase of the cortical alpha rhythm. [57, p. 71]
In their discussion of the genesis of rhythmic oscillations in the CNS, Thatcher and John [57] present evidence for alpha activity being generated at various levels of the CNS from "relatively discrete locus-to-locus thalamocortical interactions" to the cortical-wide synchronization of the EEG characteristic of "resting alpha". At all of these levels, however, thalamic circuits have an integral role. They characterize the thalamus as "a master synchronizer involved in determining information distribution and time parsing of information flow." [p. 70] The heart of this "master synchronizer" is the nucleus reticularis, whose inhibitory interneurons generate the synchronous bursts of activity that modulate both local and global distributions of alpha activity. Herkenham [37] and Steriade & Llinas [54] review more recent evidence confirming NR's "pacemaker" role. Corticothalamic pathways can effect oscillatory activity via their influence upon NR. The firing rhythms of reciprocal thalamocortical projections can be suppressed or enhanced by converging cortical (PC, PfC) or MRF influences. And more generalized effects upon rhythmic oscillations can be induced by NR stimulation of intralaminar (and paralaminar) projections to the tangential intracellular fields. But all of this requires the participation of NR.
What controls the patterns of these rhythmic waves scanning periodically across the cortex? According to the present model, it could vary from a single ensemble of modular units in visual cortex, to a broad coalition of specialized processors distributed across the entire expanse of the cortex. In either case, it is NR that modulates their combined effects upon the attentional matrix. In instances where voluntary search for a target stimulus is required, a coalition of processors in PfC would gain control of scanning; while a startled response to a loud noise would rest control of the attentional matrix away from any one group of processors, instructing them all to "pay attention" to (i.e process) whatever has just occurred (MRF).
In this scenario, "global messages" would be the functional analog of these waves traveling, tangentially, across the entire cortex. Global broadcasting would correspond to the diffuse activation of the cortex by the extended reticular-thalamic activating system, the source of these scanning waves. Global broadcasting does not add any informational content to what is represented in the cortex. Rather, it conveys higher level attentional and contextual "information" to all of the modular units of the cortex, related to the momentary focus of conscious awareness. The content of messages is implicit and globally integrative rather than explicit and differentiated.
Functionally characterized, global messages can vary from the basic command to "Pay attention to that loud noise to your right rear!" (MRF) to subtle and highly detailed attentional information. An example of the latter would be the instructions to "Tell the experimenter whenever you see a green T in the stimulus array". In this case the global message has its origins in an explicit verbal instruction. But for this instruction to guide attentional search, it must be translated by the brain into an attentional message and globally broadcast, so that all specialized processors are primed to respond, and the appropriate ensemble of them can activate to guide the subject's search strategy.
It is this "translation" process for which NR is postulated to be is essential. It serves as the nexus for both the global attentional matrix and the reticular core, generating highly coherent, correlated patterns of EEG activity distributed across the cortex via the tangential intracortical network. The holistic framework for neurocognitive processes presented in this paper differs markedly from the modular, "society of mind" paradigm which is enjoying current popularity in the cognitive and neurosciences. For this reason, and others noted earlier, much of the evidence upon which the NGW model is based is peripheral to the information processing approaches that guide most contemporary research. Many of the early research findings we cite have not been subsequently developed, and many of the more recent ones have yet to be replicated. This is not surprising given the nascent state of our knowledge of conscious processes. Much work still needs to be accomplished before the competing theories put forward over the past few years can be empirically sorted out. Our hope is that the framework offered here might contribute to a convergence of perspectives serving to guide ongoing research into the neural basis of conscious experience.
It seems appropriate to conclude with a brief comment on a possible experimental approach for testing the model. The basic premise of the NGW model is that oscillatory phenomenon play a central role in the global integration of neural systems mediating conscious processes. We have previously cited research (see pp. 17-19) implicating the tangential cortical network (TIN) in the integrated representation of global features of a visual stimulus via local, synchronous oscillations. This has been demonstrated, in primary visual cortex, for cortical columns separated by as much as 7 mm. The researchers report, however, that electrode sites greater than 7 mm apart show no such oscillatory linkages (coherency matching).
How, then, might such linkages be made among widely separated cortical columns selectively responding to, say, the orientation, color and movement of the same object, as well as the sound its makes (e.g. a humming top)? Are there, indeed, "higher-order" oscillatory systems that accomplish these more global integrations? An experiment to test the NGW model could involve something as simple as presenting two commonly employed ERP stimuli, such as a light flash and click, simultaneously, and observing response patterns from arrays of electrodes placed in primary visual and auditory cortex. We would not predict that oscillations in such widely separated areas would be tightly phase-locked. Rather, if a thalamically-generated "scanning wave" does play a role in integrating perception across modalities, there would be a predictable latency corresponding to the time it takes the wave of activation to travel between the two areas (and speed of scanning should vary as a function of arousal levels). Even for a simple, bimodal stimulus like a flash/click, latencies greater than 60 ms might be expected for conscious processing [70-72]. It would be optimal, of course, to have electrodes recording from appropriate sites in the reticular nucleus at the same time. The research cited earlier (p. 16), however, suggests that recordings from the superficial and deep layers of the cortex could provide evidence for such rhythmically integrative oscillations (layer IV neurons do no show "stimulus-linked" oscillations).
Finally, we would recommend, with Stryker [33], that researchers who wish to pursue these questions in greater depth, reacquaint themselves with the rich literature of the past 50 years "exploring the rhythms of the brain revered by the pioneers of electroencephalography", no longer dismissing such research as "irrelevant to neural information processing".
ACKNOWLEDGMENTS
We would like to thank David Galin, James Skinner and Charles Yingling for the valuable insights they have provided during the extended process of developing the ideas offered here. Our thanks to Roy John; Christopher Koch; Theodore Melnechuk and Vernon Mountcastle for reading and commenting upon earlier versions of the model. Special thanks to Francis Crick and Arnold Scheibel for ongoing dialogue and criticism concerning this area of intense mutual interest.
ABBREVIATIONS
NGW - Neural Global Workspace. This is the acronym for the explicit neural model presented in this paper. The anatomi- cal substrate of NGW is the Tangential Intracortical Network (TIN). When the term "NGW System" is used, it refers to NGW in combination with the "attentional matrix" including: NR, ILC, PC, PfC and MRF (see below).
ILC - Intralaminar Complex. The ILC includes all of the intra- laminar nuclei associated with the nonspecific thalamus. The model reviews recent evidence that "paralaminar nuclei" (Herkenham [37]) may form part of this nonspecific complex as well.
NR - Nucleus reticularis thalami. Also commonly referred to as the reticular nucleus of the thalamus.
TIN - Tangential Intracortical Network. A cortical neuropil, localized mostly to the upper layers and extending, tangentially, across the columnar architecture of the entire cortical mantle. It is formed primarily by the axon collaterals and dendritic ramifications of the pyramidal cells, but also includes axonal ramifications from cortico-cortical association fibers, including the corpus callosum.
MRF - Midbrain Reticular Formation. In the model, the primary nuclei of MRF are the superior colliculi and cuneiform nucleus. MRF, with its orderly projections upon NR (and ILC), is part of the global attentional matrix.
PC - "Posterior Cortex". PC is one of three broadly defined attentional systems of the "global attentional matrix" controlling access to the NGW. It consists of networks of corticothalamic neurons which give off collaterals as they pass through NR on their way to the principal nuclei of the thalamus. Reciprocal fibers projecting from the principal thalamus back to PC which give off collaterals to NR are also a part of this sub-system.
PfC - "Pre-frontal Cortex". PfC is one of the attentional systems of the global attentional matrix. It consists of neurons which project directly to NR, allowing pre- frontal cortex to selectively "gate" irrelevant inputs from the thalamus that would otherwise reach the posterior cortex. This sub-system provides the basis the regulation of attention in terms of goals and cognitive schemas. It may also provide the basis for working memory [65].
REFERENCES
[1] See, for example, M.S. Gazzaniga: The Social Brain: Discover ing the Networks of the Mind (Basic Books, New York, 1985).
[2] See, for example, M. Minsky: The Society of Mind (Simon and Schuster, New York, 1986).
[3] See, for example, J.L. McClelland, D.E. Rummelhart, and the PDP Research Group: Parallel Distributed Processing (MIT Press, Cambridge, MA, 1986).
[4] B.J. Baars: A Cognitive Theory of Consciousness (Cambridge University Press, Cambridge, MA, 1988). B.J. Baars: How does a serial, integrated and very limited stream of consciousness emerge out of a nervous system that is mostly unconscious, distributed, parallel and of enormous capacity? In Experimen- tal and Theoretical Studies of Consciousness, ed. by G.R. Bock and J. Marsh (John Wiley & Sons, Chichster, England, 1993).
[5] See, for example, D.C. Dennett: Consciousness Explained (Little, Brown and Co., Boston, 1991).
[6] See, for example, V.B. Mountcastle: An organizing principle for cerebral function: The unit module and the distributed system. In The Mindful Brain, ed. by G.M. Edelman and V.B. Mountcastle (MIT Press, Cambridge, 1978) pp. 21-42.
[7] See, for example, P.S. Goldman-Rakic: Changing concepts of cortical connectivity: Parallel distributed cortical networks. In Neurobiology of the Neocortex, ed. by P. Rakic and W. Singer (S. Bernhard, Dahlem Konferenzen, John Wiley, Berlin, 1988) pp.177-202.
[8] B.J. Baars: Conscious contents provide the nervous system with coherent global information. In Consciousness and Self- Regulation, ed. by R. Davidson, G. Schwartz and D. Shapiro (Plenum Press, New York, 1983) pp. 45-76.
[9] R. Jackendorf: Consciousness and the Computational Mind (MIT Press, Cambridge, 1987).
[10] P.N. Johnson-Laird: The Computer and the Mind (Harvard University Press, Cambridge, 1988).
[11] G.M. Edelman: The Remembered Present: A Biological Theory of Consciousness (Basic Books, New York, 1989).
[12] A.R. Damasio: Time-locked multiregional retroactivation: A systems level proposal for the neural substrates of recall and recognition. Cognition 33, 25 (1989).
[13] F. Crick and C. Koch: Towards a neurobiological theory of consciousness. Semin. Neurosci. 2:263 (1990).
[14] F. Crick and C. Koch: Some reflections on visual awareness. Cold Spring Harb. Symp. on Quantit. Biol. 15, 953 (1990).
[15] B.J. Baars: A Global Workspace theory of conscious experience. B.J. Baars and J. Newman: A neurobiological interpretation of a Global Workspace theory of consciousness. In Consciousness in Philosophy and Cognitive Neuroscience, ed. by A. Revonsuo and M. Kamppinen (Erlbaum, Hillsdale, N.J., in press).
[16] R. Reddy and A. Newell: Knowledge and its representation in a speech understanding system. In Knowledge and Cognition, ed. by L.W. Gregg (L. Erlbaum Associates, Potomac, MD, 1974).
[17] L.D. Erman and V.R. Lesser: A multilevel organization for problem solving using many diverse, cooperating sources of knowledge. In Proc. 4th Ann. Joint Computer Conf. (Georgia, USSR, 1975) pp. 483.
[18] B. Hays-Roth: A blackboard architecture for control. Artifi cial Intelligence 26, 251 (1985).
[19] J.L. McClelland: The programmable blackboard model of reading. In Parallel Distributed Processing, Vol. 2, ed. by J.L. McClelland and D.E. Rummelhart (MIT Press, Cambridge, 1986) pp. 122.
[20] J.F. Kihlstrom: The cognitive unconscious. Science 237, 285 (1987).
[21] J. Szentagothai: The local neuronal apparatus of the cere bral cortex. In Cerebral Correlates of Conscious Experience, ed. by P.A. Buser and A.Rougeul-Buser (INSERM Symp. No. 6 N. Holland/Elsevier, Amsterdam, 1978) pp. 131.
[22] For a review, see E.G. Jones: What are the local circuits? In Neurobiology of the Neocortex, ed. by P. Rakic and W. Singer (S. Bernhard. Dahlem Konferenzen, John Wiley, Berlin, 1988) pp. 137.
[23] W. James: The Principles of Psychology (Dover Publications, New York, 1890).
[24] J.P. Changeux and S. Dehaene: Neuronal models of cognitive functions. Cognition 33, 64 (1989).
[25] For an historical review, see D.L. LaBerge: William James symposium: Attention. Psychol. Science 1(3), 156 (1990).
[26] K.M. Heilman, R.T. Watson and E. V. Valenstein: Neglect and related disorders (Chp. 10). In Clinical Neuropsychology, ed. by K.M. Heilman and E.V. Valenstein (Oxford University Press, New York, 1985).
[27] S.A. Hillyard and T.W.Picton: Electrophysiology of cogni tion. In Handbook of Physiol. Sec. 1: Neurophysiol. Vol. V, Higher Functions of the Brain, ed. by F. Blum (Am. Physiol. Soc., 1987) pp. 519.
[28] M.I. Posner: Structures and functions of selective atten tion. In Clinical Neuropsychology and Brain Function: Re search, Measurement and Practice, ed. by T. Boll and B. Bryant (American Psychological Association, Washington, D.C., 1987) pp. 169.
[29] M.I. Posner, A.W. Inhoff, F.J. Friederich and A. Cohen: Isolating attentional systems: A cognitive-anatomical analysis. Psychobiol. 15(2), 107 (1987).
[30] For a review, see M. Mesulam: Principles of Behavioral Neurology (F.A. Davis, Philadelphia, 1985).
[31] D. Dennett and M. Kinsbourne: Time and the observer: The where and when of consciousness in the brain. Behavioral and Brain Sciences 15(2) (1992).
[32] C. von der Malsburg and W. Schneider: A neural cocktail- party processor. Biol. Cybern. 54, 29 (1986).
[33] M.P. Stryker: Is grandmother an oscillation? Nature 338, 297 (1989).
[34] F. Crick: Function of the thalamic reticular complex: The searchlight hypothesis. Proc. Natl .Acad. Sci. USA 81, 4586. (1984).
[35] H.H. Jasper: Unspecific thalamocortical relations. In Hand book of Neurophysiology, II, ed. by J. Field, H.W. Magoun and V.E. Hall (American Physiological Society, Washington, D.C., 1960) pp. 1307.
[36] For a review, see E.G. Jones: The Thalamus (Plenum Press, New York, 1985).
[37] M. Herkenham: New perspectives on the organization and evolution of the nonspecific thalamocortical projections. In Cerebral Cortex, Vol. 5, ed. by E.G. Jones and A. Peters (Plenum Press, New York, 1986) pp. 403.
[38] G. Moruzzi and H.W. Magoun: Brain stem reticular formation and activation of the EEG. Electroencephalogr. Clin. Neurophysiol. 1, 455 (1949).
[39] M.E. Scheibel and A.B. Scheibel: The organization of the nucleus reticularis: A golgi study. Brain Res. 1, 43 (1966).
[40] J.E. Skinner and C.D. Yingling: Central gating mechanisms that regulate event-related potentials and behavior. In Progress in Clinical Neurophysiology: Attention, Voluntary Contraction and Event-Related Cerebral Potentials, Vol. 1, ed. by J.E. Desmedt (Karger, Basel, 1977) pp. 30.
[41] For an historical review, see R.F. Thompson: Foundations of Physiological Psychology (Harper and Row, New York, 1967).
[42] M.E. Scheibel and A.B. Scheibel: Structural substrate for integrative patterns in the brain stem reticular core. In Reticular Formation of the Brain, ed. by H.H. Jasper, L.D. Proctor, R.S. Knight, W.C. Noshay and R.T. Costello (Little, Brown and Co., Boston, 1958) pp. 31.
[43] A. Brodal: Anatomical points of view on the alleged morpho logical basis of consciousness. Acta Neurol. 12(2), 166 (1964).
[44] G. Macchi and M. Bentivoglio: The thalamic intralaminar nuclei and the cerebral cortex. In Cerebral Cortex, Vol. 5, ed. by E.G. Jones and A. Peters (Plenum Press, New York, 1986). pp. 355.
[45] K.H. Pribram: Languages of the Brain (Brooks/Cole, Monterey CA, 1971).
[46] J.D. Green, D.S. Maxwell, W.J. Schindler and C. Stumpf: EEG "theta" rhythm: Its anatomical source and relation to activ- ity in single neurons. J. Neurophysiol. 23, 403 (1960).
[47] W.R. Adey, D.O Walter, and C.E Hendrix: Computer techniques in correlation and spectral analysis of cerebral slow waves during discriminative behavior. Exp. Neurol. 3, 501 (1961).
[48] F.O. Schmitt, P. Dev and B.H. Smith: Electrotonic process ing of information by brain cells. Science, 193, 114 (1976).
[49] Z.F. Kisvarday, K.A.C. Martin, T.F. Freund, Z.S. Magloczky, D. Whitteridge and P. Somogyi: Synaptic targets of HRP- filled layer III pyramidal cells in the cat striate cortex. Exp. Brain Res., 64, 541 (1986).
[50] C. von der Malsburg and W. Singer: Principles of cortical network organization. In Neurobiology of the Neocortex, ed. by P. Rakic and W. Singer (S. Bernhard. Dahlem Konferenzen, John Wiley, Berlin, 1988) pp. 69.
[51] E.G. Jones and A. Peters: Cerebral Cortex, Sensory Motor Areas and Aspects of Cortical Connectivity, Vol. 5 (Plenum Press, New York, 1986)
[52] R.W. Sperry: Lateral specialization in the surgically separated hemispheres. In The Neurosciences Third Study Program, Chp. 1, ed. by F.O. Schmitt and F.G. Worden (MIT Press, Cambridge, 1974) pp. 5.
[53] W.J.H. Nauta and M. Fiertag: Fundamental Neuroanatomy (Free man and Co., New York, 1986).
[54] For a recent review, see M. Steriade and R.R. Llinas: The functional states of the thalamus and the associated neuro- nal interplay. Physiol. Rev. 68(3), 649 (1988).
[55] C.M. Gray, P. Konig, A.K. Engel, and W. Singer: Oscillatory responses in cat visual cortex exhibit inter-columnar syn- chronization which reflects global stimulus properties. Nature 338, 334. C.M. Gray, A.K. Engel, P. Konig and W. Singer: Stimulus dependent neuronal oscillations in cat visual cortex: Receptive field properties and feature dependence. Eur. J. Neurosci. 2, 607 (1990).
[56] For a recent review, see S. Zeki: The visual image in mind and brain. Sci. Am. 267(3), 68 (1992).
[57] R.W. Thatcher, and E.R. John: Foundations of Cognitive Pro cesses (Lawrence Erlbaum, New York, 1977).
[58] A.B. Scheibel: Anatomical and physiological substrates of arousal. In The Reticular Formation Revisited, ed. by H.A. Hobson and M.A.B. Brazier (Raven Press, New York, 1980) pp. 55.
[59] J. Schlag and M. Waszak: Characteristics of unit responses in nucleus reticularis thalami. Brain Res. 21, 286 (1970).
[60] J. Schlag and M. Waszak: Electrophysiological properties of units of the thalamic reticular complex. Exp. Neurol. 32, 79 (1971).
[61] M. Waszak: Firing patterns of neurons in rostral and ventral part of nucleus reticularis thalami during EEG spindles. Exp. Neurol. 43, 38 (1974).
[62] C.D. Yingling and J.E. Skinner: Regulation of unit activity in nucleus reticularis thalami by the mesencephalic reticular formation and frontal granular cortex. Electroencephalogr. Clin. Neurophysiol., 39, 635 (1975).
[63] See, for example, A.R. Luria: The Working Brain (Basic Books, New York, 1973).
[64] See, for example, J.M. Fuster: The Prefrontal Cortex (Ra ven Press, New York, 1980)
[65] P.S. Goldman-Rakic: The prefrontal contribution to working memory and conscious experience. In The Brain and Conscious Experience, ed. by O. Creuzfeld and J. Eccles (Pontifical Academy Press, Rome, 1988).
[66] F.E Bloom: What is the role of general activating systems in cortical function? In Neurobiology of the Neocortex, ed. by P. Rakic and W. Singer (S. Bernhard Dahlem Konferenzen, John Wiley, Berlin, 1988) pp. 407.
[67] H.W. Magoun: The Waking Brain (Thomas, Springfield, IL, 1963).
[68] V.B. Mountcastle: Brain mechanisms for directed attention. J. Royal Soc. of Med. 71 14 (1978).
[69] S. Sharpless and H.H. Jasper: Habituation of the arousal reaction. Brain 79, 655 (1956).
[70] R.T. Knight: Electrophysiology in behavioral medicine. In Principles of Behavioral Neurology, ed. by M. Mesulam (F.A. Davis Co., Philadelphia, 1985) pp. 327.
[71] B. Libet: Neuronal vs. subjective timing for a conscious sensory experience. In Cerebral Correlates of Conscious Experience, ed. by P.A. Buser and A. Rougeul-Buser (INSERM Symp. No. 6., N. Holland/Elsevier, Amsterdam, 1978) pp. 69.
[72] B. Libet: Timing of cerebral processes relative to concom- itant conscious experience in man. In Advances in Physiolog ical Science, ed. by G. Adam, I. Meszaros and E.I. Banyai (Pergamon Press, Elmsford, NY, 1981) pp. 313.
[73] U. Neisser: Cognitive Psychology (Prentice-Hall, Englewood Cliffs, NJ, 1967).
[74] For review articles, see L.R. Squire and N. Butters: Neuropsy chology of Memory (Guilford Press, New York, 1984).