This paper explores a remarkable convergence of ideas and evidence, previously presented in separate places by its authors. That convergence has now become so persuasive that we believe we are working within substantially the same broad framework. Taylor’s mathematical papers on neuronal systems involved in consciousness dovetail well with work by Newman and Baars on the thalamocortical system, suggesting a brain mechanism much like the global workspace architecture developed by Baars (see references below). This architecture is relational, in the sense that it continuously mediates the interaction of input with memory. While our approaches overlap in a number of ways, each of us tends to focus on different areas of detail. What is most striking, and we believe significant, is the extent of consensus, which we believe to be consistent with other contemporary approaches by Weiskrantz, Gray, Crick and Koch, Edelman, Gazzaniga, Newell and colleagues, Posner, Baddeley, and a number of others. We suggest that cognitive neuroscience is moving toward a shared understanding of consciousness in the brain.
Here we outline some elements of this emerging framework.
1. Consciousness is an architectural aspect of the organization of the brain. It has global influences and effects.
“Global Workspace theory” developed by Baars (1983, 1988, 1996) has three fundamental constructs, which are most easily cast in terms of a theater metaphor. Consciousness in this metaphor corresponds to the bright spot on stage, while unconscious systems operate in the dark backstage and in the audience. Thus we have
1.1 The active bright spot on the stage of working memory;
1.2 Contextual systems which reside backstage, shaping the contents of consciousness without themselves entering consciousness.
1.3 The audience of specialized knowledge sources that are unconscious, ranging from long term memories to motor automatisms. This audience is sitting in the dark, receiving information from the bright spot on stage.
Only the global workspace is conscious at any given moment; contextual elements backstage are unconscious, as are the specialized systems of the brain that can receive information from the bright spot in the darkened theater.
2. A number
of brain mechanisms could serve such functions.
2.1 The active bright spot may be closely associated with sensory projection areas which provide the detailed contents of sensory consciousness, closely supported by subcortical systems in the thalamus, basal ganglia, hippocampus and brain stem. An attentional system works to select which of several potential inputs will become conscious at any given time. The reticular nucleus of the thalamus (nRt), with its ability to gate traffic from sensory surfaces to cortex, is one natural candidate for such a selective attention mechanisms (Newman & Baars, 1993; Baars, 1993, 1996; Posner (1994) has found evidence for several other brain loci for visual attention, notably the anterior cingulate, which may be associated with effortful attention.
2.2 Contextual systems which are unconscious but still shape conscious contents may be found, among other places, in the dorsal visual pathway, including right parietal cortex, which provides the object-centered “framework” for visual perception. (See Baars, 1996)
2.3 The theater audience of specialized knowledge sources that receive conscious input would include frontal cortex, memory-related systems such as the hippocampus, brain mechanisms for automatic routines such as the basal ganglia, and those mediating emotional processes such as the limbic system and amygdala.
2.4 How could visual conscious contents be distributed to their audience? The brain’s gross anatomy shows massive fiber tracts leading from posterior to anterior cortex, from one hemisphere to the other, and from cortex to thalamus and back again. Each of these massive fiber tracts is known to map significant features of conscious information from one location to another. For example, the famous work on split-brain patients by Sperry, Gazzaniga and Bogen shows that cutting the transverse fiber tracts between the hemispheres blocks the normal flow of consciously available information between the right and left hemispheres. Thus broadcasting of cortical sensory information may occur in several directions from sensory cortex: from back to front, from side to side, and from cortex to midbrain and back again.
Baars (1988) hypothesized the neural substrate of this GW to be an “Extended Reticular Thalamic Activating System” (ERTAS). The term ERTAS emphasizes that this system (while including the classical Reticular Activating System), is #centered upon the thalamus# with its rich, reciprocal projections with the cortex. Newman and Baars (1993) also cite the significance of Layer I of the cortex, often viewed as a “feltwork” of neurons that permit not just axonal transmission, but horizontal spread of activation as well. Layer I may also provide a channel for global broadcasting; however, it seems better suited to connectionist propagation than to high-fidelity cortico-cortical transmission. In more recent developments, cortical sensory projection areas seem more plausible loci of a conscious global workspace, at least for sensory consciousness.
3. Conscious experience involves the ceaseless interaction of input (from outside and within the brain) with memory.
In the vocabulary suggested by Taylor, it is relational: "The conscious content of a mental experience is determined by the evocation and intermingling of suitable past memories evoked by the input giving rise to that experience." Such a relational feature implies a ceaseless dialogue between conscious contents and unconscious systems, notably the hippocampus, basal ganglia, and unconscious regions of cortex.
3.1 The process of 'intermingling' requires both long-ranged inter-modal and short ranged intra-modal competition to support the intermingling process. Taylor has explored the problem of inter-modal (global) inhibition via the nucleus reticularis thalami (nRt) (Taylor, 1992). The nRt seems a beautiful device to achieve global input competition, being composed of solely inhibitory cells and also having remarkable lateral connectivity similar to the outer plexiform layer of the retina (Taylor 1990). Taylor suggests the phrase "Conscious I" for the nRt contribution to global competition (Taylor, 1993), with the notion that nRt may bind activities across modalities. He believes that nRt supports a far more complex form of spatial competition, however, than given by a simple winner-take-all (WTA) form, since it produces as winner a wave over the whole of the coupled thalmic-nRt-cortical system (Alavi and Taylor, 1993, 1995). There will be many spatial regions over cortex which have non-zero activity in such a winner.
3.2 In a recent target article for *Brain and Behavioral Science*, Jeffrey Gray (1995) reviewed a wealth of data supporting a model for "the generation of the contents of consciousness ... correspond[ing] to the outputs of a comparator that, on a moment-by-moment basis, compares the current state of the organism's perceptual world with a predicted state" (p. 659). The heart of this "comparator system" is the hippocampus (see Figure 8); but its activities are closely tied to those of the basal ganglia and cortex as well
Newman (1995b) offered the alternative hypothesis that Gray's comparator provides feedback to the ERTAS, "flagging" present perceptions as "expected/familiar" or "unexpected/novel". A novel/unexpected flag interrupts ongoing behavior, causing an ERTAS-mediated orienting response. Conversely, an expected/familiar flag would produce habituation or, in the case of goal-directed behavior, a shift in attention to the next step in the current program.
In his response to the BBS commentaries, Gray (1995) offered an extension
to his original model "that links the outputs of the comparator system
to the reticulo-thalamic core which, as set out by Newman, seems likely
to underlie the generation of consciousness-as-such. (p.716)" He described
findings by Lavin & Grace (1994) showing that the fornix also projects
to areas in the basal ganglia, that project to the nucleus reticularis
thalami (nRt). The possible role of the nRt in the selection of attention
and conscious processing has been pointed out by Crick (1984) and has been
incorporated into a neural-network model by Taylor (1992). Note that, since
the output to these neurons is itself inhibitory, its activation has the
effect of disinhibiting these sensory relay pathways, that is, increasing
the entry to the cerebral cortex of those stimuli that are currently engaging
the thalamocortical loops. (Gray, 1995, p. 712)
4. Cortical foci of activity appear to provide the contents of consciousness.
Taylor suggests that the crucial aspect for the emergence of consciousness form nonconscious neural activity is the creation of relatively long-lasting 'bubbles' of activity in cortex by local recurrence (Amari, 1977). It has been proposed (Taylor, 1996b) that such activity singles out those cortical regions which have highest density of cells in layers 2 and 3, which appear to be identical to the sites of highest coding in cortex. This model then allows some of the crucial aspects of qualia (transparency, ineffability, intrinsicality) to be derived (Taylor, 1996b), as well as giving a neural underpinning to the 'global workspace' idea (Baars, 1988) by means of the excellent intercortical connectivity of the main such regions in parietal, inferotemporal and frontal cortices. Thus the detailed coding of the content of consciousness would thereby appear to be cortical. How much the thalamus is involved in this coding is still to be determined. This view seems to be in complete accord with the extension of the ERTAS system (Newman and Baars, 1993) and helps flesh it out further.
4. There is competition for access to consciousness at any given time between different sensory inputs, including “inner” sensory inputs like visual imagery and inner speech. Other possible inputs include abstract conscious contents such as beliefs and ideas.
What is selected to be conscious involves both top-down influences and bottom-up inputs. Top-down influences include goal systems associated with frontal cortex, and emotional and motivational systems combining limbic and right frontal cortex. Thus conscious contents reflect an endless interplay of competition and cooperation between possible inputs and top-down influences.
Taylor points out that there is also a further totally inhibitory system of great importance in its contribution to global competition, this being the basal ganglia. These are expected to be able to support various of the processes of comparison and buffering needed in the development of sequence learning (schemata) and of rule formation and reasoning; all of these are involved in anterior cortex (Ca). The analysis and lodelling of this system has been explored in Bapi et al (1996) in which a full discussion of a neural model of executive function of the frontal lobes is developed. There has also been exploration of the psychological aspects of such an architecture for temporal tasks (such as delayed matching tasks) (Monchi and Taylor, 1996). The principles of such an architecture have also been explored as part of the ACTION network (Taylor, 1995; Taylor and Alavi, 1993).
5. The brain stem-thalamocortical axis supports the state, but not the detailed contents of consciousness, which are produced by cortex.
Newman & Baars (1993) described the anatomy, and putative functions, of a "Neural Global Workspace" in detail. Reviewing half-a-century of research, the paper described an architecture extending from the midbrain reticular formation to prefrontal cortex. The midbrain core of this system is the classic Ascending Reticular Activating System, including the intralaminar complex (ILC). Integrated with this reticular core is a "global attentional matrix", centered upon nucleus reticularis (nRt), which both 'gates' information flow to and from the cortex, and regulates electrical oscillatory patterns throughout it. (Footnote 1)
When Newman & Baars (1993) was published the authors were unaware of
Taylor's model for "global guidance", mediated by nRt. Newman refers to
this particular form of resource allocation as "global attention". Its
core neural mechanism is an _array of gating circuits_ contained in the
*nucleus reticularis thalami* (*nRt*), which covers the lateral surfaces
of the thalamus. Through it pass nearly all the pathways coursing between
the thalamus and cerebral hemispheres. Via these reciprocal nRt circuits,
the cerebral cortex and brainstem reticular formation _selectively modulate
their own information processing activities_ in the service of conscious
perceptions, intentions and/or plans.
6. A selective attention system, including the reticular nucleus of the thalamus (nRt), selects among multiple possible conscious contents.
Attentional selection of conscious contents appears to operate by means of the reticular nucleus of the thalamus (nRt), under the influence of several centers from brain stem to prefrontal cortex. The nRt is believed to “gate” the sensory thalamic nuclei, which flow up to the primary projection areas of the cortex, and is therefore located in the most strategic position for controlling sensory input. It is known that almost ten times as many visual neurons go down from visual cortex to the thalamus as go the other way, suggesting continuous and intimate looping between thalamus and cortex, controlled by nRt.
A voluntary component may also act via the anterior cingulate gyrus and the vigilance network of prefrontal cortex, identified by Posner (1994). Attentional selection is also influenced by the contents of consciousness, suggesting that re-entrant loops from sensory cortex descend to the major sensory nuclei of the thalamus to create a self-sustaining feedback loop, again under control of nRt (Edelman, 1989; Newman, 1995 ab).
7. Posterior cortex provides sensory conscious contents, while anterior cortex is involved in active, voluntary control; both seem to be necessary for normal conscious experience and control.
From Taylor’s point of view, it is possible to discern at least two main components of consciousness,which may be denoted by Cp and Ca. The former of these, Cp, is that part of consciousness containing phenomenal experience, the 'raw feels'. The subscript 'p' denotes the passive nature of Cp, without activetransformations being made by frontal motor cortical systems such as involved in thinking and reasoning. The subscript can also be used to denote posterior, in terms of the siting of this component in posterior cortical systems and their related thalamic structures. The relevant memorial structures needed to give content to Cp are those of the primary sensory, postprimary, associative and heteromodal cortices, all of which are posterior.
There is furthermore the antithetical Ca, in which 'a' denotes active as involved in thinking and planning as well as responding in a willful manner. It has its major activity involved with anterior cortical, related thalamic and basal ganglia neural sites, which are well known to be crucially involved with actions of either direct (motor action) or higher order (thought, planning) form. In Baars’ vocabulary the emphasis would be on the voluntary character of such frontal activity. A more complete analysis is presented in Taylor (in preparation).
As Newman writes,
The aspect of consciousness most completely controlled by these parallel distributed circuits is that of immediate perception, mainly supported by posterior cortical modules, allowing them to be interpreted as a passive part of the overall complex making up the totality of conscious experience. It is nRt which allows arrays of modular processors to compete for access to the cortical GW. This competition is global due to the effects of dendro-dendritic connections among nRt neurons which create a series of coherent, nRt-wide activation/inhibition patterns. It is this nRt coherence which mediates the global allocation of the processing resources of this posterior corticothalamic system
8. Interaction of self structures and consciousness.
Baars (1988) develops a framework in which one necessary condition for consciousness is providing information to a self-system, viewed as a stable contextual system that provides predictability across a variety of situations. The left-hemisphere “narrative interpreter” discovered by Gazzaniga and colleagues may be one such self-system.
Taylor has suggested further components such as a self structure of consciousness, Cs, and an emotional part Ce. The former of these is undoubtedly present and sited in the meso-orbital frontal region (as lesion studies of frontal patients indicate, starting with the famous case of Phineas Gage); it is unclear that the latter has any specific cortical site but is more diffusely distributed from limbic systems, especially by the dopaminergic and related aminergic neurotransmitter systems.
9. Layers of visual consciousness.
Blindsight is one of the empirical sources of any approach to visual consciousness. It seems to show that the first visual projection area of the cortex, the first place where the visual tract reaches cortex, is a necessary condition for conscious vision, but not for unconscious visual knowledge. Blindsight patients, who lack parts of this area (V1) have unconscious (implicit) visual knowledge about objects, such as location, color, motion and even object identity. But they passionately protest that they have no conscious visual experience in the damaged parts of the visual field.
Crick and Koch (1995) have pointed out a significant paradox in blindsight: Area V1 is known to represent points (center-surround), and spatial intervals (spatial frequencies), but it does not represent color, motion, or visual form. Yet these higher-level visual features drop from visual consciousness when damage occurs to V1, as shown by blindsight patients. In Crick and Koch’s language, V1 does not “explicitly” represent color and form, because it has no cells that are sensitive to color, etc.. Yet the conscious appreciation of color and form is destroyed by damage to V1. We will call this “the paradox of area V1.” An adequate theory of visual consciousness must account for this paradox.
In contrast to the puzzle of V1, damage to higher visual areas of cortex creates only selective loss of conscious features. Thus damage to area V4 destroys conscious perception of motion, but not of color, form, or retinal locus. Damage to area IT (the lower temporal cortex) may destroy conscious object recognition, but not color, retinal locus, and motion. And so on. From that point of view V1 is especially important because its absence serves to abolish all visual conscious features, including those not explicitly represented in V1.
It is important to take into account the remarkable interconnectivity of all parts of the brain. In the visual system there is top-down feedback from each “higher” visual area to lower areas, as well a bottom-up flow going the other way. Indeed, in the case of thalamocortical loops, the down-flowing neurons outnumber the up-going ones by a ratio of almost ten to one. This makes sense in terms of classical neural net approaches, in that multiple layers, when their patterns of activation are consistent, serve to enhance each other, while nonconsistent layers tend to compete and decrease activation patterns. This result has been called the “rich get richer” principle (McClelland and Rumelhart, 1986).
How does visual consciousness relate to other kinds of sensory and nonsensory consciousness? There is evidence for blindsight analogues in auditory and somatosensory cortex. These may work much like the early visual areas. One can think therefore of the early sensory projection areas as several theater stages, each with its own bright spot, alternating rapidly in such a way that at any given moment only spotlight is on. As the most active, coherently paced sensory cortex becomes conscious, its contents are broadcast widely throughout the brain, simultaneously suppressing the other sensory cortices for a cycle time of approximately 100 msec., the time needed for perceptual integration. The most highly amplified sensory projection area in any 100 msec. time cycle may send conscious information to spatial, self, and motor maps via massive fiber tracts, some going through thalamus, others via the corpus callosum, and yet others through massive cortico-cortical connections within each hemisphere. Thus rapid switching can take place between different sensory cortices, integrated by multimodal neurons in association cortex and elsewhere.
Any theory of consciousness must satisfy a considerable set of empirical constraints that are quite well-established (Baars, 1983, 1988, 1996). Together, these sources of evidence implicate a Global Workspace system. which is most easily understood via the metaphor of the theater of consciousness. In this metaphor consciousness corresponds to the bright spot on the stage of a theater, shaped by context systems that are invisible behind the scences, and being broadcast to a set of receiving processors that can be viewed as the unconscious audience. Taylor has emphasized the need to include a memory-matching component, which he calls “relational,” leading to a Relational Global Workspace framework. We believe this framework captures our current understanding of the psychological and brain evidence.