SINGLE-NEURON THEORY OF CONSCIOUSNESS
Journal of Theoretical Biology (2005, In Press)
Steven Sevush
Departments
of Psychiatry and Neurology
University
of Miami School of Medicine
1400 NW 10
Ave, Suite 702
Miami,
Florida 33136, USA
ssevush@med.miami.edu
Tel:
305-243-4082
This work
was presented in abstract form at
"Toward
a Science of Consciousness" held in Tucson, April 2002.
ABSTRACT:
By most accounts, the mind arises from
the integrated activity of large populations of neurons distributed across
multiple brain regions. A contrasting
model is presented in the present paper that places the mind/brain interface
not at the whole brain level but at the level of single neurons. Specifically, it is proposed that each
neuron in the nervous system is independently conscious, with conscious content
corresponding to the spatial pattern of a portion of that neuron's dendritic
electrical activity. For most neurons,
such as those in the hypothalamus or posterior sensory cortices, the conscious
activity would be assumed to be simple and unable to directly affect the
organism's macroscopic conscious behavior.
For a subpopulation of layer 5 pyramidal neurons in the lateral
prefrontal cortices, however, an arrangement is proposed to be present such
that, at any given moment: i) the
spatial pattern of electrical activity in a portion of the dendritic tree of
each neuron in the subpopulation individually manifests a complexity and
diversity sufficient to account for the complexity and diversity of conscious
experience; ii) the dendritic trees of the neurons in the subpopulation all
contain similar spatial electrical patterns; iii) the spatial electrical
pattern in the dendritic tree of each neuron interacts nonlinearly with the
remaining ambient dendritic electrical activity to determine the neuron's
overall axonal response; iv) the dendritic spatial pattern is reexpressed at
the population level by the spatial pattern exhibited by a synchronously firing
subgroup of the conscious neurons, thereby providing a mechanism by which
conscious activity at the neuronal level can influence overall behavior. The resulting scheme is one in which
conscious behavior appears to be the product of a single macroscopic mind, but
is actually the integrated output of a chorus of minds, each associated with a
different neuron.
Key Words: binding problem; consciousness; hard problem; neural correlate of consciousness; single-neuron theory.
1.
INTRODUCTION
The purpose of this paper is to suggest a
shift in emphasis from the large to the small in the search for a brain
correlate for the mind. Currently, the
general view is that the mind arises from the integrated activity of large
populations of neurons distributed across multiple cortical and subcortical
brain regions (Sperry 1969; Crick and Koch 1990a; Llinas et al. 1998; Damasio
1999, Edelman and Tononi 2000, John 2001, Singer 2001). Dynamic mechanisms, such as those involving
reentrant flow of information (Edelman and Tononi 200), synchronous electrical
oscillatory activity (Singer 1998; Engel et al. 1999; Sauve 1999; Sewards and
Sewards 2001), and top-down attentional effects (Reynolds and Desimone 1999;
Wolfe and Cave 1999), have then been invoked to provide for the
"binding" of the dispersed neuronal activity into a unified stream of
consciousness. The perspective,
according to this view, is that the activity within any single neuron
correlates with merely a fragment of the total conscious experience; it is only
through the integration of these fragments that a single whole-brain
consciousness is assumed to emerge.
A contrasting model is outlined in the
present paper that places the mind/brain interface not at the whole brain level
but at the level of the single neuron.
According to the model, a single brain at any given moment harbors many
separate conscious minds, each being associated with the activity of a
different neuron. For most neurons,
only a simple conscious experience is assumed to be present. For a subpopulation of neurons populating
the lateral prefrontal cortices, however, a complex consciousness is assumed to
be separately present in each neuron, with each of the neurons having a similar
conscious experience, and with that conscious experience being that which in
most models is attributed to the joint action of vast numbers of neurons
distributed throughout the brain. The
theory posits a mechanism that explains how the separately conscious neurons
might express their output in terms of whole brain behavior. The resulting scheme is one in which
conscious behavior, while appearing to be the product of a single macroscopic
mind, is actually the result of the assembled output of a chorus of minds, each
associated with a different individual neuron.
The main features of the theory can be
summarized as follows:
a) Each neuron in the nervous system is
assumed to be independently conscious, with a component of the electrical
activity in the neuron's dendritic tree serving as the neural correlate of
consciousness (NCC) for that neuron.
b) For most neurons, such as those in the
hypothalamus or those in the posterior sensory cortices, or for cortical
interneurons, the conscious activity of the neuron is assumed to be simple and
unable to directly affect the organism's macroscopic behavior. Such neurons would not, therefore,
contribute to what is usually taken as a person’s conscious behavior.
c) For layer 5 pyramidal neurons in the
lateral prefrontal cortices, however, the electrical activity within a portion of
the apical dendritic tree is presumed to be by itself complex and diverse
enough to account for the complexity and diversity of conscious experiences
usually ascribed to the activity of the brain as a whole. Additionally, as a result of the diversity of
afferent input that uniquely characterizes the lateral prefrontal cortices, the
layer 5 pyramidal neurons in these areas can plausibly be regarded as
recipients of input pertaining to all of the sensations, thoughts, feelings,
and memories that make up conscious experience.
d) A mechanism can be delineated by which
a subgroup of synchronously firing layer 5 lateral PFC neurons with similar
conscious experiences can express that experience by means of the spatial
pattern they manifest at the neuronal population level. The axonal output of these neurons then
transmits the conscious content to other brain regions, including those that
control behavior.
e) As a result, while the content of
conscious experience is seen to correlate with activity in individual neurons,
the production of conscious behavior results from population activity at the
network level.
The most immediate implication of this
arrangement is that it offers a novel approach to solving the "binding
problem," the problem of explaining how spatially dispersed neuronal
activity can correlate with an apparently seamless single experience. The currently popular views are that
either: a) perceptual unity is achieved
through temporal synchrony of the component elements, perhaps via 40 Hz gamma
frequency oscillations (Singer 1998; Engel et al. 1999; Sauve 1999; Sewards and
Sewards 2001); or b) perceptual unity is an illusion, with only coordinated
behavioral output being in need of an explanation (Dennett 1991; Shadlen and
Movshon 1999). These approaches have
their difficulties, however. The
evidence for temporal synchrony as a basis for binding has been criticized on
both technical and conceptual grounds (Hardcastle 1994, 1997; Gold 1999;
Revonsuo 1999; Bieberich 2002; Edwards 2005), while the argument that
perceptual unity is an illusion has been challenged both empirically and
philosophically (Chalmers 1996; Robertson 2003; Bayne and Chalmers 2003). The theory presented in this paper proposes
an alternative explanation for perceptual binding, with perceptual unity being
achieved through spatial convergence of incoming signals upon single neurons.
In what follows I present an outline of
the single-neuron model and argue for its feasibility given what we know of
brain architecture and function at both the macroscopic and microscopic
levels. I will then compare the model
with other popular theories of consciousness and describe the empirical
predictions made by the model that might adjudicate between the competing
perspectives. I will conclude by
offering a brief speculation as to how the model might be extended to an
ultramicroscopic domain, and how this extension might be used to formulate a
framework for addressing the philosophically difficult aspects of the
mind/brain problem.
The model, it should be stressed, is at
this point presented as a hypothesis, not a proven theory. The contention, given the data presently
available, is not that the proposed single-neuron mechanism is necessarily the
way the brain processes information, only that it is a possible way for it to
accomplish the task and that it should not be dismissed out of hand. There is no question that the theory takes a
stance that is at odds with prevailing intuitions. Yet, the stakes are high and, in view of the difficulties
inherent in the mind/brain problem, unusual approaches to its solution may be
warranted.
2.
TERMINOLOGY AND THE POSSIBILITY OF MULTIPLE CONSCIOUSNESSES
With regard to terminology, when the
terms "mind" and "consciousness" appear in this paper, they
are being used in the specific sense that Chalmers uses the term
"experience" in his discussion of the easy and hard problems of
consciousness (Chalmers 1996). As
described by Chalmers, the "easy" problems of consciousness are those
that appear tractable by the usual methods of science and include the ways in
which the brain focuses attention, integrates information, controls behavior,
and so on. In contrast, the
"hard" problem of consciousness is the problem of accounting for the
content of subjective experience, that elusive datum that appears to accompany
certain high level brain processes. Why
does the electrical processing in the brain that ensues when one looks at the
sky give rise to the experience of "blue" and not to the experience
of some other color, or to no color experience at all? In what follows, it will be to consciousness
in this sense that a neural correlate will be sought.
An additional matter is the potential ambiguity
associated with the proposal that multiple minds are present in a single brain,
a feature that plays centrally in the proposed theory at both the macroscopic
and microscopic levels. The problem is
that since multiple consciousnesses are assumed to be present in a single
brain, and since these include consciousnesses of varying types, how will we
know what is to be denoted by the term "consciousness?" That is, to which of the different kinds of consciousness will the term refer?
In order to minimize the potential for
ambiguity of this kind, the theory will initially be developed for the
restricted case of verbally reportable consciousness (designated in this paper
as "VR-consciousness"), by which I mean consciousness that has direct
access to verbal report. It must be
stressed that this restriction is a temporary one, and will be introduced only
as a convenience to minimize ambiguity in the initial presentation of the
theory. It is not implied that
VR-consciousness is the only type of consciousness present in the brain, nor
that VR-consciousness is even the only high level consciousness present in the
brain. Nor is it implied that verbal
output is the only form of behavior that can be influenced by VR-consciousness. VR-consciousness has been chosen for this
purpose because it is around this form of consciousness that the debate over
the hard problem of consciousness has been primarily focused. Once the theory has been examined for the
special case of VR-consciousness, the argument will then be expanded to account
for the presence of other types of consciousness.
Armed with these preliminaries, we can
now turn to the presentation of the model.
A sketch of the single-neuron theory for VR-consciousness will be
developed in two steps, one macroscopic, the other microscopic. At both levels, a divergent/convergent
feedforward model of information flow (Abeles 1991) will be adopted. In the macroscopic step, the suggestion will
be made that within a divergent/convergent feedforward framework, much of the
distributed brain processing usually taken as the correlate of VR-consciousness
can be recast as either previous or subsequent to VR-conscious experience. Plausibility arguments will be offered in support
of the contention that only when the information flow reaches the left[1]
lateral prefrontal cortex (PFC) does it achieve VR-conscious status. In the microscopic step, the argument will
be made that it is not the left lateral PFC as a whole that mediates
VR-consciousness, but rather that VR-consciousness is mediated separately and
redundantly by individual pyramidal neurons inhabiting the region. A divergent/convergent feedforward model of
information flow involving a subset of individual neuron within the population
will be offered. Finally, the theory
will be extended to account for consciousness of a nonverbally reportable
variety.
3. THE MACROSCOPIC STEP: PLAUSIBILITY OF A FOCAL MODEL
The first objective in developing the
single-neuron theory is to determine, on a macroscopic level, just which brain
region or regions participate directly in VR-conscious experience. The commonly held view is that it is the
activity of neuronal populations dispersed across multiple cortical and
subcortical brain regions that mediates VR-consciousness (Llinas et al. 1998;
Damasio 1999, Edelman and Tononi 2000, John 2001). Importance is usually placed not only on cortical processing but
also on thalamocortical loops (Llinas et al. 1998; Edelman and Tononi 2000) and
involvement of the upper brainstem and other subcortical structures (Damasio
1999). With regard to cortical
information flow, while there is general acknowledgment of a hierarchical
arrangement for posterior cortical subregions (Hubel and Wiesel 1962, 1965;
Rockland and Pandya 1979; Maunsell and Essen 1983; Felleman and Van Essen 1991;
Barbas and Rempel-Clower 1997; Lamme and Roelfsema 2000; Inui et al. 2004), the
flow of information between posterior and frontal cortex is usually regarded as
inherently bidirectional and lacking any feedforward directional bias (Fuster 1998).
On the face of it, this multifocal view
of conscious information processing seems unavoidable. It is indeed difficult to imagine how a
single brain region could serve as a focus of convergent inputs sufficiently
rich to account for the manifest intricacy and multifarious nature of
VR-conscious experience. Additionally,
the multifocal view is encouraged by data obtained from brain ablation,
neuroimaging, and single-cell electrical recording studies, each of which has
demonstrated the importance of widely dispersed brain regions in even the
simplest VR-conscious tasks. Thus, the
ablation paradigm, which regards a brain region as important for a given task
only if damage to that brain region regularly results in the loss of ability to
perform the task, has implicated brainstem, thalamic, limbic, and widespread
cortical regions in VR-conscious processing (Crick and Koch 1990a; Damasio
1998; Parvizi and Damasio 2001). More
recently, neuroimaging studies employing functional MRI and related imaging
techniques or with those utilizing single-cell recordings from pyramidal
neurons in various brain areas have provided direct evidence of a role for
widespread cortical and subcortical regions in VR-consciously mediated behavior
(Crick and Koch 1990a; Dehaene et al. 1998; Llinas et al. 1998; Raichle 2000).
Despite these arguments, it remains
possible that while a distributed model is consistent with the available data,
so might be a convergent model. It may
be possible that just a subset or even only one of the areas delineated in the
above studies actually subserves VR-conscious experience and that the other
areas function only to provide information to this convergent region. For example, during a VR-conscious visual task,
ablation, neuroimaging, and single-cell recording data have together implicated
activity spread across primary and secondary occipital cortices, tertiary
temporal and parietal association cortices, PFC, and multiple diencephalic and
brainstem structures in the context of VR-conscious experience. Since, however, these disparate regions are
highly interconnected, the possibility that only one of the identified regions
actually subserves VR-conscious experience and that the other regions serve
only to provide input to this focal region cannot be dismissed without further consideration.
For a focal model to gain serious
consideration, a brain region would need to be identified that could serve as a
locus of convergence for the full set of stimuli that feed VR-conscious
experience. Such a region should
manifest certain empirically demonstrable features. First, the area should be the recipient of afferent connections
from brain localities corresponding to each of the sensory, emotional,
mnemonic, and other components that make up VR-conscious experience. Second, it should be the case that the area
is adequately connected with brain regions capable of mediating VR-conscious
behavioral output. Third, the area
should show activation with neuroimaging or single-cell recording techniques
during engagement in VR-consciously mediated tasks. And fourth, ablation of the area should result in loss of
VR-conscious experience. In the
following, the argument will be made that the left lateral PFC (corresponding
to portions of Brodmann areas 9, 10, 11, 46, 47) appears to satisfy each of the
enumerated requirements.
Afferent Connections: The first question is whether the left
lateral PFC is appropriately positioned to be the recipient of afferent
connections pertaining to each of the sensory, emotional, and mnemonic
components that comprise VR-conscious experience. This possibility is often dismissed summarily, but invariably without an accompanying in-depth analysis (eg, see Dehaene
1998). Yet a review of the
neuroanatomical literature suggests that the idea of a convergence zone should
not be so casually disregarded. It has
been suggested, for example,
that the PFC as a whole functions as a convergence zone, receiving input from
most other brain regions (Miller and Cohen 2001; Elston 2003). Since PFC subregions are strongly
interconnected (Barbas and Pandya 1991), any one PFC subregion could be capable
of serving as a convergence target for all the other PFC subregions. In the single-neuron theory, it is the left
lateral PFC that is assumed to serve as the final convergence area for the
sensory, emotional, and mnemonic components of VR-consciousness, with evidence in support of this
contention available for each of these components.
With regard to sensory input, the lateral
PFC has long been established to receive input from the occipital, temporal,
and parietal association cortices that process the incoming visual, acoustic,
and tactile signals that provide the organism with information about its extrapersonal
and proprioceptive space (Kuypers et al. 1965; Pandya and Kuypers 1969;
Goldman-Rakic and Schwartz 1982;
Petrides and Pandya 1984, 1999; Seltzer and Pandya 1989; Barbas and
Pandya 1989; Pandya and Yeterian 1990), including simultaneous convergent input
from multiple modalities (Jones and Powell 1970; Chavis and Pandya 1976; Bruce
et al. 1981). Additionally, the lateral
PFC receives inputs from the caudal orbitofrontal cortical regions that
function as association areas for taste and smell (Baylis et al. 1995, Johnson
et al. 2000), with inputs for the latter bypassing the thalamus and projecting
solely to the frontal lobes.
With regard to emotion, ablation and
activation studies have each supported a role for both anterior cingulate and ventromedial
orbitofrontal cortices in assessing the emotional significance of stimuli
(Baleydier and Mauguiere 1980, Bush et al. 2000, Phan et al. 2002). Anatomically, these regions are the
recipients of input from the amygdala, the hypothalamus, and the hippocampus
(Porrino et al. 1981; Amaral and Price 1984; Goldman-Rakic et al. 1984; Barbas
and DE Olmos 1990) which carry the processed
visceral stimuli from which emotions are proposed to be comprised
(Damasio 1999). Both regions project
strongly to lateral PFC (Eslinger and Damasio 1985) where, it is suggested
(Damasio 1999; Gray et al. 2002), emotion and cognition become integrated, with
feelings providing “value,” which is needed for decision-making.
With regard to memory, links to lateral
PFC are well established for both short-term working memory and long-term
distraction-stable memory. For
short-term working memory, findings from ablation, neuroimaging, and
single-cell recording studies point to the lateral PFC as a principal site
(along with temporal and parietal tertiary association cortex) wherein
information is retained for brief periods during the performance of complex
problem-solving tasks (Fuster and Alexander 1971; Kubota and Niki 1971;
Goldman-Rakic 1992; Cohen et al. 1997; Courtney et al. 1997; D'Esposito et al.
1998; Rainer et al. 1998, 1999; Romo et al. 1999; Fletcher and Henson 2001;
Constantinidis and Goldman-Rakic 2001; Constantinidis et al. 2002).
For long-term distraction-stable memory,
the lateral PFC plays a key role in the retrieval of episodic memories stored
in posterior neocortex and hippocampus (Goldman-Rakic 1992; Fuster 1999), a
component of memory that is inherently conscious (Tulving 2002). Direct connections between the hippocampus
and the lateral prefrontal cortex have been established, and ablation (Wheeler
1995) and neuroimaging studies (Fletcher and Henson 2001; Braver et al. 2001,
Slotnick et al. 2003) have provided direct evidence of a relationship between
lateral PFC and long-term memory function.
Efferent Connections: On the efferent side, the question is
whether the left lateral PFC is able to access the motor programs that
presumably mediate VR-conscious behavioral output. There is ample evidence that this is the case for lateral PFC,
which sends outputs to the frontal eye fields, the premotor frontal cortex, the
basal ganglia, cerebellum, and superior colliculus (Goldman and Nauta 1976;
Alexander et al. 1986, Bates and Goldman-Rakic 1993; Lu et al. 1994; Schmahmann
and Pandya 1997). For VR-consciousness
in particular, a critical issue is whether the left lateral PFC is connected
adequately with Broca's area, the brain region that provides for syntactical
linguistic output. The pivotal role of
Broca's area, situated in the posterior part of the inferior frontal gyrus
(Brodmann's areas 44 and 45) of the left frontal lobe, in language expression
was established first in 1861 (Broca 1861) and has been amply confirmed in the
one and a half centuries that have followed.
For logical, grammatical output, which is the type involved in verbally
reporting on VR-conscious experience, the situation is asymmetric, with only
the "dominant" hemisphere (the left hemisphere in most humans and
almost all right-handers) capable of performing the function. Left lateral PFC is ideally positioned to
influence Broca's area and thereby express verbal reports pertaining to
VR-conscious experience since it lies immediately adjacent to it and projects
to it strongly (Deacon 1992).
Activation Studies: The principal evidence for activation of the
left lateral PFC during VR-conscious tasks comes from studies of working
memory, a cognitive function that is usually regarded as VR-conscious (Baddeley 2003). As noted above, neuroimaging and single-cell
recording studies indicate that lateral PFC is a principal mediating site for
working memory. Additionally, left
lateral PFC has been shown directly to become active in neuroimaging studies
during tasks involving VR-conscious experience (McIntosh et al. 1999; Kjaer et
al. 2001, Stephan et al. 2002).
Event-related potential measurements in humans have shown further that
lateral PFC is activated subsequent to posterior cortex in association with
performance of VR-conscious tasks, with a mean latency of activity in response
to visual stimuli over occipital areas measured at 56 msec and over lateral PFC
at 80 msec (Foxe and Simpson 2002).
This temporal sequence of events would be in keeping with a model in
which posterior cortex provides preliminary processing of stimuli for subsequent
VR-conscious processing by lateral PFC.
Results of Ablation: If the left lateral PFC were, by itself, the
mediator of VR-conscious experience, then ablation of the region should result
in its elimination. Care is needed, however,
in applying this syllogism in clinical situations. To begin with, since the focus up to this point has been
specifically on VR-consciousness, not on higher consciousness in general,
clinical cases would need to be assessed for the presence or absence of this
particular form of consciousness if they were to serve as an appropriate test
of the proposed theory as presented so far.
As we will see when we extend the discussion of the theory beyond the
confines of consciousness of the verbally reportable variety, it would be
damage to the lateral PFC bilaterally that would be expected to eliminate
higher consciousness in general.
There is also a methodological concern
having to do with the extent of the cortical damage incurred in reported cases. As will be evident when the theory is described on the microscopic
level, only nearly complete
inactivation of the left lateral PFC would be predicted to eliminate
VR-consciousness, and only complete inactivation of the lateral PFC bilaterally
would be predicted to eliminate higher consciousness more generally. Thus,
we note that some authors (Bogen 1995; Alexander and Stuss 2000; Taylor
2001) have cited clinical examples in the literature, such as the celebrated
case of Phineas Gage (Damasio 1999), where
extensive prefrontal lesions failed to eliminate higher conscious experience,
and have used these to argue
that the lateral PFC by itself cannot be the sole mediator of higher
consciousness. As pointed out by Crick
and Koch (1998), however, in none of these cases did the damage include the
entirety of the lateral PFC, rendering their significance uncertain. With only partial removal of the relevant
cortex, the remaining tissue might have been able to maintain residual higher
conscious function. On the other hand,
situations in which the lateral PFC is known to have been entirely destroyed
bilaterally, such as in patients with advanced Pick's disease, massive
bilateral strokes, extensive anoxic damage, or massive surgical bilateral
frontal lobe ablations as reported by Walter Dandy (1946), the clinical picture
is of a persistent vegetative state in which the victim is devoid of any higher
level consciousness (Sato et al. 1989; Laureys et al. 1999; Heilman and
Valenstein 1993; Laureys 2004). And in
cases in which the left lateral PFC is known for certain to have been destroyed
entirely, such as in global aphasia from a massive left middle cerebral artery
stroke, VR-consciousness is clearly lost, with the right hemisphere providing
residual, nonverbal higher conscious function (Heilman and Valenstein
1993). While such cases are consistent
with the notion that VR-consciousness is mediated solely by the left lateral
PFC, and that higher consciousness in general is mediated solely by the lateral
PFC bilaterally, the extensive nature of the damage in these reports ruins the
specificity of the correlation.
Cases are needed in which the left lateral PFC (or the lateral PFC
bilaterally) has been completely removed or rendered inoperative while other
structures have been left
intact, and an assessment made specifically for the presence or absence of
VR-consciousness (or higher consciousness more generally). Such cases have yet to be reported.
In summary, anatomical and functional
studies have been reviewed that point to the importance of left lateral PFC for
VR-conscious experience. Other authors
have also emphasized the importance of the lateral PFC to consciousness (Crick
and Koch 1998; Rees et al. 2002; Stephan et al. 2002). The present theory goes further, however, in
proposing that the left lateral PFC, situated within a divergent/convergent
feedforward information processing system, serves by itself as the direct
neural correlate for VR-consciousness.
On the other hand, suggestions have occasionally been made favoring
other brain regions as a center for consciousness (Dandy 1946; Taylor
2001). The parietal tertiary
association cortex is an attractive possibility (Taylor 2001), but the parietal
lobes do not receive direct information pertaining to smell, which is a common
component of VR-conscious experience.
Alternatively, the orbitofrontal cortex, which is a region that receives
convergent input from all modalities, might be considered to be a candidate for
a localized NCC. However, complete orbitofrontal
ablations, while they impair emotional processing, do not eliminate
VR-consciousness (Eslinger and Damasio 1985; Heilman and Valenstein 1993;
Damasio 1999).
The localization to the left lateral PFC
proposed in this paper, it should be noted, may not be the full extent to which
VR-consciousness is localized. The
lateral PFC might itself be further divisible.
In fact, support has been garnered for the lateral PFC encompassing two
functionally distinct subregions, one dorsolateral (portions of Brodmann 9, 46)
the other ventrolateral (portions of Brodmann 10, 11, 47), the two subregions
differing with respect to both anatomical connections and type of information
being processed (Oliveri et al. 2001). It has been suggested that the established
division of the posterior cortex into "dorsal" and
"ventral" streams (Ungerleider and Mishkin 1982; Goodale and Milner
1992; Milner and Goodale 1993; Foxe and Simpson 2002) continues into the
frontal lobes, with the dorsolateral PFC receiving input from the dorsal stream
via the parietal association cortex and the ventrolateral PFC receiving input
from the ventral stream via the temporal association cortex (Milner and Goodale
1995). Since both areas are strongly
connected to Broca's area, both are in a position to affect verbal output. Further, since both areas are strongly interconnected, either could
ultimately be the sole mediator of VR-conscious experience with the one serving to provide input to or receive output from the other. In fact, Milner and Goodale (1995) have argued for just this possibility,
suggesting that only the ventral stream is associated with VR-conscious
experience and that the dorsal stream processes information VR-unconsciously.
In the discussion that follows, the question
of which subregion by itself might mediate VR-conscious experience will be left
undecided and the generic term "lateral PFC" will be used to refer to
the area in general. In any case,
whatever the final localization of the macroscopic anatomical substrate for
VR-consciousness might be, all that is needed for the further development of
the single-neuron theory is that a macroscopically focal correlate for
VR-consciousness exists that is relatively homogeneous and is in a position to
receive all the sensory, emotional, and mnemonic stimuli that characterize the
VR-conscious experience.
4.
THE MICROSCOPIC STEP: THE SINGLE-NEURON
THEORY
Subsequent to the question of which
macroscopic brain region mediates VR-consciousness is the question of how the neurons
populating this region accomplish the task.
According to the generally accepted view, the participating neurons join
their activity into a single VR-conscious experience via dynamic interactive
mechanisms. The single-neuron theory
posits instead that a subpopulation of neurons in the left lateral PFC, rather
than joining their activity into a single VR-consciousness, remain separately
VR-conscious and produce the illusion of a joint VR-consciousness by virtue of
the joint action of their outputs. In
the model, each of the VR-conscious neurons at any moment is proposed to
experience not merely a fragment of a higher-order joint VR-consciousness but,
rather, the whole range and complexity of VR-conscious experience generally
attributed to the population as a whole.
Plausible candidates for
the role of VR-conscious neurons are the layer 5 pyramidal cells of the left
lateral PFC. A pivotal role for these
neurons in conscious processing has been suggested by other investigators
(Crick 1994; Orpwood 1994; Gilbert 1998; Bieberich 2002). It will be useful to briefly review the main
characteristics of these neurons before describing how they might individually
serve as neural correlates of VR-consciousness.
Anatomical
and Functional Characteristics of Lateral PFC Layer 5 Pyramidal Neurons
Like all neurons, layer 5 pyramidal cells
function by integrating the electrical responses of an enormous array of
dendritic synaptic inputs, the result
of which determines a single channel of axonal output consisting of
sequences of action potentials of identical shape and amplitude. Early on, the dendritic tree was modeled as
a single entry point (McCullough and Pitts 1943), with inputs simply summing
linearly to produce axonal output. In
the late 1950s, Rall (1959, 1967) introduced the passive cable model, in which
the spatial extent of the dendritic tree was taken into account and
accommodation was made for the effects of location and amplitude of
postsynaptic potentials on the neuron's axonal response. The passive cable model provided for a
degree of intradendritic information processing that was not possible with the
earlier "point neuron" model.
In recent years, the presence of significant nonlinear intersynaptic
effects occurring on multiple levels has been discovered (Cash and
Yuste 1999; Magee
1999; Koch 1999; Larkum et al. 1999a; Magee and Cook 2000; Spruston 2000;
Hausser et al. 2001; Wei et al. 2001), suggesting that considerable intradendritic computational processing
precedes the eventual generation of axonal action potential.
Additionally,
there is evidence that intradendritic processing occurs independently in
separated compartments before an overall effect is delivered to the soma (Koch
1997). For layer 5 pyramidal neurons, a
unique anatomy facilitates this compartmentalization. Most notable is the long ascending dendritic shaft, extending 1.3
mm from the cell body in layer 5 to the apical tip in layer 1, that separates
the proximal basal dendrites from the distal apical dendrites. Midway along the shaft are the oblique
dendrites, both proximal and distal.
Recent evidence suggests that the output of the various compartments
interact via two intraneuronal dendritic action potentials. The first is a forward propagating,
calcium-mediated potential originating at the distal end of the dendritic stalk
that is produced as a result of nonlinear integration of apical dendritic
inputs. This action potential courses
from the apical end of the dendritic shaft to the soma where it induces a burst
of axonal action potentials (Helmchen et al. 1999; Schiller et al. 1995, 1997; Schwindt
and Crill 1999; Larkum et al. 1999a; Williams and Stuart 2002). The second is a backpropagating
sodium-mediated action potential that is triggered by axonal firing and that
courses from the soma back into each of the dendritic arbors (Stuart and Sakmann
1994). The two action potentials appear
to be coupled, with the backpropagating potential lowering the threshold for
triggering of the forward propagating potential, provided the two potentials
occur within a few milliseconds of each other (Larkum et al.1999a, 1999b; Larkum et al. 2001; Schaefer et al.
2003; Larkum et al. 2004). The coupling has been shown to depend
crucially on the morphology of the participating dendritic branches (Mainen and Sejnowski 1996; Schaefer et al. 2003; Vetter et al. 2001).
Aside from the nonlinear interactions
present at the intercompartmental level, nonlinear interactions have been found
within the dendritic arbors as well (Cash and Yuste 1999; Magee 1999; Koch 1999; Larkum et
al. 1999a; Magee and Cook 2000; Spruston 2000; Hausser et al. 2001; Wei et
al. 2001; see Magee
2000 for a review). It has been
suggested that nonlinear interactions between neighboring synapses provide for
increased computational capability in dendritic trees (Segev and London 2000;
Van Ooyen et al. 2002; see Hausser and Mel 2000 for a review).
Because of the anatomy and
compartmentalization of the layer 5 pyramidal neurons, they are ideally
positioned to integrate input from the different cortical layers (Spratling
2002; Thomsen et al. 2002a; Briggs and Callaway 2005).
Thus, the basal dendrites, which are situated near the cell body in
layer 5, receive intralaminar input from other pyramidal neurons in the same
layer (Lubke et al. 1996; Thomson et al. 2002b; Douglas and Martin 2004). Higher up, the oblique dendrites receive
inputs from layers 2, 3 and 4 (Schaefer et al. 2003). Still more distally, the apical dendrites receive inputs from
layers 1, 2 and 3 (Schaefer et al. 2003).
A functional interpretation of these
inputs is provided by a proposed microscopic feedforward circuit that includes
the layer 5 pyramidal neurons (Gilbert and Wiesel 1986; Thomson et al. 2002a,b;
Thomson and Bannister 2003). The main
pathway in this circuit is characterized by a feedforward flow beginning with
the layer 4 stellate (granular) neurons, which receive incoming feedforward
signals from other cortical and thalamic regions, and moving on to the
pyramidal neurons in layers 2 and 3, then on to layer 5 pyramidal neurons, and
finally out to pyramidal neurons in layers 1 and 5 elsewhere in the prefrontal
cortex, as well as to premotor cortex, posterior cortex, thalamus, basal ganglia,
and superior colliculus. Return input
from these distant targets arrive via diffusely projecting layer 1 fibers,
whose signals are received by the pyramidal neurons in layers 2 and 3, and 5
(the latter receiving inputs from layer 1 both directly and also indirectly via
inputs from layers 2 and 3). In this
scheme, the input from layer 1 consists of diffusely encoded signals derived
from thalamus, hippocampus, and posterior and frontal limbic areas (Barbas and
Rempel-Clower 1997; Rempel-Clower and Barbas 2000; Mitchell and Cauller 2001;
Barbas and Hilgetag 2002). These inputs
have been postulated to be associated with attention, feedback, and context
(Williams and Stuart 2002; Douglas and Martin 2004)). The inputs to layer 2 and
3 pyramidal neurons derive from both local and distant layer 2 and 3 pyramidal
neurons located throughout the lateral PFC.
These layer 2 and 3 pyramidal neurons are, especially in the lateral
PFC, highly interconnected and have been proposed to serve an important role in
the maintenance of reverberatory circuits underlying working memory (Kritzer
and Goldman-Rakic 1995; Gonzalez-Burgos et al. 2000; Melchitzky et al.
2001).
An important implication of these pathways
and connections (see Larkum 2001), and one that will be used in developing the
single-neuron model, is that the apical dendrites of layer 5 pyramidal neurons,
by virtue of being recipients of layer 2/3 pyramidal cell output, are in a
position to receive inputs representing the full range of VR-conscious content. In the following sections, we will build on
this and the other characteristics of layer 5 pyramidal neurons to present a
specific way in which the single-neuron theory of consciousness might play out
at the microscopic level. It should be
noted that it is not contended that the model to be presented is the only way a
single-neuron theory can be instantiated.
For example, it may be that neurons other than the layer 5 pyramidals
might equally well serve as loci for VR-consciousness, or it may turn out that
the details of layer 5 pyramidal function as described above need
modification. The purpose of the
presentation is mainly to provide a specific instantiation of the theory in
order to demonstrate that a single-neuron model is in principle feasible. The model will be built upon a framework of
convergent/divergent feedforward information flow, in which feedback and
reentrant circuits play only an adjunctive role. The features of the model will be presented in a feedforward
sequence: A) first, a mechanism of information encoding at the neuronal
population level will be described for the network of neurons that will serve
as the afferent input for VR-conscious neurons; B) second, the process by which signals converge upon the
dendritic trees of VR-conscious neurons and induce intradendritic NCCs will be
considered; C) third, a
mechanism by which the dendritic NCCs reexpress themselves at the neuronal
population level will be suggested.
A) Population Encoding of the
Information to be Inputted to the VR-Conscious Neurons:
We begin by examining the
information processing characteristics of layer 2 and 3 pyramidal neurons in
left lateral PFC, which will be presumed to be the source of the inputs to layer 5 pyramidal neurons that
lead to VR-consciousness. In overview, it is hypothesized that the
convergent input to the left PFC as a whole, argued in the previous section to
contain the sensory, emotional, and mnemonic signals that feed VR-consciousness,
passes from layer 4 stellate neurons to layer 2/3 pyramidal neurons, and that
the relevant information is encoded in the spatial arrangement of a
synchronously firing subpopulation of these layer 2/3 neurons.
We now consider this process in more detail.
As a preliminary, we need
first to address the nature of the neural code. There has been an ongoing debate over how information is encoded
by neuronal populations in neocortex.
The principal questions have to do with the relative importance of
individual neuronal spikes in cortical processing and whether spikes derived
from different neurons are bound by means of temporal synchrony.
The early view was that information is encoded in the cortex by the
average rate of axonal action potential production over a given time interval,
typically between 20 and 200 msec, and that the exact timing of individual
spikes is unimportant. (Adrian 1926; Barlow 1972; Shadlen and Newsome 1994;
Shadlen and Newsome 1998; Shadlen and Movshon 1999; Yazdanbakhsh et al. 2002;
Rolls et al. 2003). In recent years,
however, the alternative possibility, that each spike individually carries
relevant information, has gained a growing following (Softky and Koch 1993;
Gray 1999; Borst and Theunissen 1999; Williams and Stuart 2000; Abeles
2004). A related idea, that brain
neuronal activity can be synchronized, and that synchrony is in fact a widely
occurring and robust phenomenon that might serve to define functional
relationships between spatially distributed cortical neurons, has also gained
considerable support (Abeles
1991; Konig et al. 1995; Singer
and Gray 1995; Lumer et al. 1997; Stevens and Zador 1998; Lachaux JP et al.
1999; Usrey and Reid 1999; Grammont and Riehle 2003).
The possible mechanisms underlying the induction of temporal synchrony
of cortical neuronal activity have been explored (Niebur et al. 2002; Reyes 2003; Segev 2003; Nowotny and
Huerta 2003), as has its
possible behavioral relevance (Sougne and French 2001). In particular, it has been suggested that
synchronous firing promotes synergy between spikes arriving simultaneously at
target neurons, thereby providing a powerful signal selection mechanism (Niebur
et al. 2002; Averbeck and Lee 2004).
For the theory presented in this paper,
it will be assumed that individual spikes do matter, and that synchronous
firing identifies subpopulations of neurons that encode salient information in
their spatial arrangement. The model
employed is an adaptation of the
well-studied "synfire chain" mode of cortical signal transmission
(Abeles 1991, 2004; Aviel et al. 2002), which invokes the importance of both
individual spike timing and the synchronization of spatially dispersed neuronal
activity. The synfire chain theory
describes a process of synchronous convergent/divergent feedforward signal
transmission through multiple neuronal layers.
Synchrony is hypothesized to result from an arrangement in which, for
each layer, there is a pool of neurons with the property that each neuron in
the pool: a) receives afferent inputs
from all the neurons in the pool in the previous layer, and b) projects its
output to all the neurons in the pool in the next layer. Empirical observations supporting the
presence of synfire chain activity in the brain have been reported by a number
of investigators (Abeles 1991; Miller 1996; Prut et al. 1998; Abeles 2004;
Ikegaya et al. 2004) and their properties have been analyzed through
theoretical simulations (Abeles 1991; MacGregor et al. 1995; Aertsen et al.
1996; Hertz and Prugel-Bennett 1996; Diesmann et al. 1999; Arnoldi et al. 1999;
Gewaltig et al. 2001; Sougne and French 2001; Yazdanbakhsh et al. 2002). Theoretical simulations have indicated how,
in a multilayered feedforward neural network consisting of integrate-and-fire
neurons, synchronous firing on a millisecond time scale arises in successive
layers, even if the initial input to the chain are asynchronous (Pinsky 1995;
Arnoldi and Brauer 1996; Hertz and Prugel-Bennett 1996; Marsalek et al. 1997;
Rudd and Brown 1997; Campbell et al. 1999; Neltner et al. 2000; Niebur et al. 2002; Reyes 2003; Segev 2003; Nowotny and
Huerta 2003). The synchronous chains
have been shown to be stable against noise, provided a sufficient number of
neurons populate each pool, estimated to be on the order of 100 neurons (Abeles
1991; Postma et al. 1996; Diesmann et al. 1999; Gewaltig et al. 2001; 2001). In the stable state, essentially all response spikes in a volley
fall within ±1 ms, which is within the precision of experimentally observed
neuronal firing patterns, despite the presence of a membrane time constant of
10 ms or more (Diesmann et al. 1999; Gewaltig et al. 2001).
In the proposed single-neuron theory, the
presence of a synfire chain comprising two consecutive pools of neurons is hypothesized: one pool in lateral PFC layer 2/3, the other
pool in lateral PFC layer 5. The pool
in layer 2/3 is taken to
comprise a subpopulation of synchronously firing pyramidal neurons that have
received feedforward input from layer 4 pertaining to VR-consciousness, with
that information being encoded in the spatial arrangement of the participating
neurons. The presence of identifiable
spatial patterns of neuronal excitation in cortical neuronal populations has
been reported in a variety of contexts.
These include the peculiar complex
logarithmic relationship between spatial patterns of excitation in the
retina and those in area V1 pyramidal neurons (Schwartz 1980; Alexander et al.
2004), the specific topographic pattern of layer 5 pyramidal neurons in cat
visual cortex that exhibited repeating temporal spike sequences (Ikegaya et al
2004), the specific spatial patterns of neuronal response to air currents in
the cricket cercal system (Jacobs and Theunissen 2000), and the work of
Grossberg et al. (1999), which implicates an important role for the topography
of primate visual cortex in shaping neuronal responses at subsequent levels of
cortical visual processing. In the
single-neuron theory, it is the topography of neuronal activation in the
synchronously firing layer 2/3 pyramidal population that is assumed to carry
the information to be delivered to the layer 5 pyramidal neurons.
B) Conversion of Afferent
Population Patterns into Intradendritic Neuronal Electrical Patterns:
The next issue in the feedforward
progression is the matter of how the afferent signals from the synchronously
firing layer 2/3 pyramidal neurons converge upon layer 5 pyramidal neurons and
lead to intradendritic electrical activity that serves as the neural correlate
for VR-consciousness. Consideration
will be given first to the synaptic activation patterns induced in the layer 5
pyramidal neurons, and then to the conversion of these synaptic activation
patterns into spatial electrical patterns within the dendrites. For the model to work, two hypotheses are
needed:
HYPOTHESIS 1: THE MAPPING FROM POPULATION TO
DENDRITIC IS HOMOTOPIC AND
INFORMATION PRESERVING: Despite its
vital importance to an understanding of neuronal information processing, there
is little available data pertaining to the neurons of origin for particular
synaptic inputs within a dendritic tree.
That is, while neurons of origin have been identified with respect to
one or another overall dendritic compartment, similar information with respect
to individual synaptic inputs within a compartment are not yet available. In the absence of such empirical clues, I would suggest that a plausible possibility is that the topographic mapping of
inputs from the synchronously firing layer 2/3 pyramidal neurons onto layer 5 pyramidal
dendrites is homotopic (the shape of the one can be continuously transformed
into the shape of the other) and information preserving.
Homotopic, information preserving mappings have been found elsewhere in the nervous system, such as in the
complex logarithmic mapping of
retina onto V1 already mentioned (Schwartz 1980; Alexander et al. 2004),
and in the homotopic connections between the cerebral hemispheres via the
corpus callosum (Mitchell and Macklis 2005).
The suggestion is that the axons arriving from the synchronously firing
layer 2/3 pyramidals retain their relative positions and preserve topological information in coursing toward their
targets and do not
"criss-cross" on their way to innervating the dendritic tree of the
layer 5 pyramidals. Whether this is in
fact the case will need to be determined experimentally. Since it plays a key role in the
hypothesized theory, testing for its presence will serve as an empirical test
for the plausibility of the proposed theory.
HYPOTHESIS 2: SIMILAR PATTERNS ARE RECEIVED BY
MULTIPLE NEURONS: A second assumption is that the
projections from layers 2 and 3 to layer 5 involve convergence of a similar
pattern onto multiple neurons. If the topographic pattern of synaptic
excitation is relevant in determining whether it induces axonal firing or not,
then this is in effect the central assumption of the synfire chain
theory that was discussed in detail above, and upon which the currently
proposed model is based.
With these hypotheses in hand, a
mechanism can be described by which the spatially encoded pattern at the
population level in layer 2/3 reappears as a spatially encoded pattern of
intradendritic electrical activity in layer 5 pyramidal cells, this pattern
serving as the neural correlate for VR-consciousness. We start by proposing
that the spatial electrical pattern corresponding to VR-consciousness forms
within the thin distal branches of the apical dendritic tree. While alternative locations, such as the thin oblique dendritic branches,
could also be chosen, the apical branches are selected here to provide a
concrete example of how the overall model would work. By hypothesis 1, it is assumed that a specific pattern of
synaptic excitation is induced by convergent input from a synchronously firing
pool of layer 2/3 pyramidal neurons, and that this pattern of synaptic
excitation is a homotopic, information
preserving transform of the spatial pattern of the layer 2/3 neuronal
pool. By hypothesis 2, it is assumed
that a similar pattern of input is received by a pool of layer 5 pyramidal
cells.
The shape of the electrical response
within the dendritic branches of each neuron receiving this pattern of synaptic
input will, in turn, be an ordered transformation of the shape of the incoming
signals (Bieberich 2002, Orpwood 1994, and Livingstone 1998), modified by the
effects of synapse efficiency and local nonlinear effects (including those due
to the state of depolarization of the synapses at the time of signal input, the
interactive effects between synapses due to induced local chemical and
electrical changes, and the morphology of the dendritic tree itself), each of
which is likely somewhat different for each of the neurons in the receiving
pool (Spruston et al. 1995; Mainen and
Sejnowski 1996; Hoffman et al. 1997; Agmon-Snir et al. 1998; Cash and Yuste
1999; Magee 1999; Koch 1999; Larkum et al. 1999a; Magee and
Cook 2000; Spruston 2000; Magee
2000; Vetter et al. 2001; Hausser et al. 2001; Wei et al.
2001;; Schaefer et al.
2003; Polsky et al. 2004).
The variations in the shape of the intradendritic electrical responses
across the different neurons would be assumed to be relatively small, so that
variations in VR-conscious experiences of the different neurons would be
relatively small as well. I will
consider later the potential impact that
large variations might have on the single-neuron theory.
Of critical importance to the
singe-neuron theory is the assumption that the intradendritic electrical
pattern that forms in the distal apical branches, which is proposed to serve as
the direct neural correlate for VR-consciousness, extends through only a
portion of the neuronal dendritic tree.
It is hypothesized that its ultimate
effect on axonal firing therefore comes only after it combines nonlinearly with other
intradendritic electrical activity, both within the more proximal apical tuft
in generating a calcium-mediated dendritic action potential and in the
subsequent interaction between the different dendritic compartments. That is, the local electrical activity that
serves as the neural correlate for VR-consciousness in the distal apical tree
of a given neuron is assumed to affect neuronal output only in the context of
other inputs received by the neuron at the same time. Again, this effect is presumed to be different for the different
layer 5 pyramidal neurons in the pool, and therefore the responses of the
different neurons to the same (more or less) VR-conscious electrical activity
forming in the distal apical dendrites will be different. In the next section, this fact will be used in proposing a mechanism by which the
intradendritic VR-conscious shape reexpresses itself at the neuronal population
level. Before turning to that
mechanism, it is important to consider whether the electrical activity
purported to serve as the neural correlate for VR-consciousness has the right
type and degree of complexity to encode
for conscious experience.
COMPLEXITY OF INPUTS TO THE LAYER 5 PYRAMIDAL
NEURONS: For the single-neuron theory to be
plausible, it needs to be the case that the electrical activity in the distal
apical branches of an individual neuron is by itself complex enough to support
the complexity of VR-conscious experience.
The relevant observation pertaining to this requirement is that cortical
pyramidal neurons receive a huge number of afferent signals over their
dendritic trees. It is estimated that
for humans each cortical pyramidal neuron incorporates between 20,000 and 40,000
synapses within its dendritic tree (Abeles 1991), a number larger than is
usually appreciated and one that allows for an impressive complexity of
information encoding. For example, if
only one tenth of one percent of the synapses of a cortical pyramidal neuron
were assumed to participate in a given VR-conscious experience, and if only a
simple binary code were being used (that is, one in which the synapse is either
active or inactive), then 240 or about one trillion different
patterns could be encoded by that neuron.
Alternatively, as noted by Bieberich
(2002), if it is assumed that 5,000 simultaneously arriving synaptic input
bytes, corresponding to 5,000 active spines in the dendritic tree, are turned
over with a frequency of 50 Hz, then a single neuron would be capable of
processing 5 mB of information per second.
This, Bieberich argues, constitutes a rate of information processing
adequate to plausibly account for the complexity of VR-conscious experience.
By either of these calculations, individual
cortical pyramidal neurons appear capable of processing information of a
complexity greater than often appreciated, and quite possibly sufficient to
mediate the entirety of the VR-conscious experience.
CONVERGENCE OF MODALITIES UPON INDIVIDUAL LAYER 5
NEURONS: Another requirement that the
single-neuron theory must meet is that the dendritic trees of the VR-conscious neurons would need to
individually be recipients of all of the kinds of inputs that make up the
VR-conscious experience. If it is
assumed that the lateral PFC as a whole receives all of the inputs that
comprise VR-conscious experience, then the question is whether it plausible
that there are individual pyramidal neurons within the region that do likewise? That is, is it plausible that the
information that converges upon the region as a whole converges separately on
each of a subgroup of pyramidal neurons populating the region? In support of the presence of such an
arrangement is the clinical observation that focal lesions within lateral PFC
do not result in modality-specific deficits but only diminish the region's
function in a general manner. This
suggests at least some degree of microscopically redundant convergence
throughout the extent of the region.
Additionally, individual cortical neurons typically receive inputs from
thousands of other neurons (Koch 1997, Schuz 1998) and so are in a position to
serve as nodes of converging information.
Aside form general considerations, however, are the reports that direct
single-cell recordings have identified neurons in lateral PFC that respond
selectively to conjoint visual and auditory stimuli (Aou et al. 1983), to
conjoint visual, auditory and tactile stimuli (Tanila et al. 1992), and to
conjoint object and location features (Rao et al. 1997). Whether a subgroup of
lateral prefrontal neurons exists in which each neuron in the subgroup receives
convergent input from all the sensory, emotional, and mnemonic stimuli that
comprise VR-conscious experience is yet to be empirically determined.
C) Generation of the Outgoing
Signal:
Supposing that the spatial pattern of
distal apical intradendritic electrical activity of left PFC layer 5 pyramidal
neurons serves as the neural correlate for VR-conscious experience, there is
then the vexing question of how the information encoded at the dendritic level
is passed on to brain output mechanisms that control behavior. In particular, how can dendritic signal
patterns of high complexity converge into a single axonal output channel
without a drastic loss of information?
On the face of it, it would appear that such information loss is
unavoidable, since axons transmit action potentials that are uniform in shape
and amplitude and therefore only the timing of the transmitted spikes can be
used to encode information (Adrian 1926).
Across a range of settings, the information contributed by individual
axonal spikes has been measured to be on the order of one bit to 3 bits per
spike, with an axonal information transmission rate therefore reaching no more than
several hundred bits per second (Bialek et al. 1991, Borst and Theunissen
1999). This rate of information
processing is far smaller than that performed by the dendritic tree. How, then, can a gross loss of information
be averted?
A possible solution to this problem may
reside in another application of synchronous neuronal firing, one that is
inspired by a model of neuronal
information processing recently proposed by Bieberich (2002). The
idea is that a mechanism by which the VR-conscious electrical activity
present within the layer 5 pyramidal dendrites might be reexpressible
at the population level in the form of
the spatial arrangement of a synchronously active subset of the layer 5
pyramidal cells. Conceptually, the sequence is as follows:
1) First, as described above, similar
VR-conscious spatial patterns of electrical activity are formed in the distal
apical trees of a population of synchronously activated layer 5 pyramidal
neurons in the left lateral PFC.
2) Next, the VR-conscious spatial patterns
interact nonlinearly, at both the intracompartmental and intercompartmental
levels, with the other inputs received by these layer 5 neurons.
3) The results of these interactions are
assumed to differ across neurons in the population and therefore to have
different effects on the propensity for axonal spike generation in the
different neurons in the population, such that some neurons fire while others
remain silent.
4) As a consequence of (3), only a subgroup
of neurons in the population respond with synchronous firing, with the active
subgroup marking out a spatial pattern at the population level. The spatial pattern, it will be noted, would in part have been dictated by the spatial pattern of the
VR-conscious dendritic electrical activity in the neurons of the population.
5) If there are N neurons in the starting
population and only M neurons fire (M<N), then the number of different
subgroups that can be formed (and therefore the number of different spatial
arrangements of firing neurons that can arise) is given by the combination of N
things taken M at a time: N!/M!(N-M)!
For even small populations, this is a huge number. For example, say there are 100 layer 5
pyramidal neurons that receive synchronous dendritic inputs of similar shape
from the layer 2/3 neurons, and that half of these 100 neurons fire. Then the number of different spatial
patterns that can form would be equal to the combination of 100 things taken 50
at a time, or (100!)/(50!)(2) = 4 x 1012, an enormous number that
matches the complexity of dendritic input as calculated above.
6) The spatial pattern formed by the
subpopulation of synchronously firing left lateral PFC layer 5 pyramidal
neurons, which incorporates the interaction between VR-conscious information
and other information received by the layer 5 pyramidal neurons, serves as the
source of conscious output from the left lateral PFC to other brain regions.
While this mechanism may provide a means
for the information contained in the intradendritic VR-conscious electrical
patterns to be re-expressed at the population level, there is the additional
question of whether enough neurons are present in the proposed synchronous
populations in layers 2/3 and 5 to affect gross behavior. As so far described, this would appear
unlikely, since previous work with synfire chains has usually associated the
phenomenon with relatively small neuronal groups (Abeles 1991). A realistic model would need to incorporate
the possible presence of many small competing synchronous neuronal groups. For a single dominant synchronous group to
arise, still additional assumptions would have to be made. This issue is, of course, not unique to the theory
presented here. It is a challenge that
must be faced by most current neuronal network models of consciousness.
A plausible approach to solving this
problem may be to invoke the concept of attention as a mechanism for amplifying
the effects of selected synchronously firing neuronal pools. Studies have, in fact, been reported that
suggest a role for attention in increasing the synchrony of neuronal firing
(Steinmetz 2000; Fries et al. 2001), possibly mediated by gamma frequency
oscillations (Crick and Koch 1990a,b; Niebur et al. 1993; Niebur and Koch
1994). In this way, attention might
serve as an "on-off" switch for lateral PFC conscious function. The detailed nature of the attentional
mechanism by which the small population of synchronously firing neurons might
gain control over VR-conscious output systems can only be speculated on at
present. One possibility is that the
"bursty" output of the layer 5 pyramidal neurons (Crick 1994; Schiller et al. 1995, 1997;
Helmchen et al. 1999; Schwindt and Crill 1999; Larkum et al. 1999a; Williams
and Stuart 2002), in association with cortico-thalamo-cortical loops (Llinas et al. 1998), may provide
for the proliferation of the synchronous groups. This could conceivably be accomplished through competition, as in
the "neural darwinism" mechanism described by Edelman (1993), or
perhaps via merging of separate synfire chains (Arnoldi et al. 1999; Hayon et
al. 2005). Yet another possibility
might involve the creation of a fractal architecture in which multiple copies
of the synfire neuronal pool arise as a result of the fractal geometry of
axonal branching (Bieberich 2002). With
each of these proposed mechanisms, the lateral PFC would be envisioned as
functioning in a “holographic” manner, with the same information represented
repeatedly in neurons dispersed across the full extent of the region. As a result, it would be expected that
damage to the entire lateral PFC would be required to eliminate the
corresponding conscious functioning, a
point that was stressed earlier in Section 3.
In summary, a case has been made for the
possibility that, at the microscopic level, VR-consciousness is mediated by
single neurons rather than by a neuronal population. The macroscopic step was necessary to provide for individual
neurons that receive all the inputs comprising VR-conscious experience and for
the supposition that there might be groups of such neurons that individually
receive similar input. In the
microscopic step, it was argued additionally that individual neurons are by
themselves sufficiently complex to account for the complexity of VR-conscious
experience and that a mechanism may be present that translates the spatial
pattern of dendritic activation within synchronously firing neurons into the
spatial arrangement formed by those neurons at the population level.
5. ADDITIONAL
CONSIDERATIONS:
Other Conscious Regions: The discussion so far has been limited to
the case of VR-consciousness, that form of consciousness largely responsible
for the debate over the "hard" problem of consciousness. It was useful to invoke this restriction in
order to narrow the focus of the presentation.
The theory can, however, be extended to include forms of consciousness
that are not verbally reportable.
Indeed, patients with lesions that result in the loss of
VR-consciousness often appear to manifest residual nonverbal consciousness,
presumably mediated by structures outside the left lateral PFC. Damasio (1998, 1999) has suggested that the
residual consciousness observed in such patients be regarded as one of two
overlapping types of consciousness, a non-verbal "core consciousness"
mediated by midline cortical and subcortical structures, and a linguistically
competent "extended consciousness" mediated by lateral neocortical
structures. The notion that a single
brain might contain multiple centers of consciousness has been suggested by
other investigators as well (Geschwind 1981, Edelman and Tononi 2000).
A particularly intriguing possible center
of nonverbal consciousness is the right lateral PFC. Since this region is
anatomically homologous to the left lateral PFC, being anatomically distinguishable from the latter principally by its
lack of direct connections with Broca's area and related structures, it might
be possible that it mediates a nonverbal form of higher consciousness separate from that mediated by the left
hemisphere. In 1973, Rolando Puccetti
suggested exactly this possibility (Puccetti 1973), proposing that the apparent
duplication of consciousness that results from corpus callosum resection (the
so-called "split-brain" procedure) is best explained by assuming that
the right and left hemispheres maintain separate conscious streams not only
after callosal disconnection but before the disconnection as well. According to his hypothesis, the callosal
section merely serves to uncouple what are already, in the normal state, two
anatomically separate consciousnesses.
Such a notion would fit well with the single-neuron theory, which
readily admits to the existence of multiple anatomically separate centers of
conscious experience.
In Puccetti's model, it should be noted,
the hypothesis is that there are no more than two conscious centers present,
and the discussion is limited to the macroscopic level only. With the presently proposed theory, an
additional step is taken, that of proposing that many macroscopic conscious
centers may be present in a single brain, and that the multiplicity of
consciousnesses extends to the microscopic single-neuron level as well. The overall picture would be one in which
multiple macroscopic centers of consciousness are proposed to exist, including
ones located throughout neocortical, subcortical, and brainstem regions, with
each center composed, in turn, of populations of neurons that individually and
redundantly mediate the conscious activity appropriate to the brain region
within which they reside.
Relationship between the Single-Neuron
Theory and Other Anatomical Theories of
Consciousness:
The theory presented in this paper is compatible with many of the
anatomical and functional features that characterize other recently proposed
theories of consciousness. For example,
Dehaene (1998) places the lateral PFC in a prominent position in his
description of Baars' global
workspace theory (Baars 1988; Baars and
Franklin 2003), Llinas (1998) focuses on thalamo-cortico-thalamic
oscillating loops, Damasio (1998) distinguishes between a brainstem mediated
core consciousness and a cortically mediated extended consciousness, Dandy
(1946) favors a striatal dominance in his theory of consciousness, and Edelman
(1993) proposes a theory based on “neural darwinism.” The neuroanatomical and functional proposals made in these
theories are for the most part compatible with the single-neuron theory. The principal differences are that with the
single-neuron theory, the lateral PFC by
itself is assumed to serve as the macroscopic correlate for higher
consciousness, with the other structures playing only an adjunctive role, and
the individual neuron is assumed to serve as the microscopic correlate for
consciousness of all kinds, while in the other models, consciousness is always
a network property.
Philosophical Caveat: The legitimacy of employing
"experience" as a target of scientific investigation is not, it
should be noted, universally accepted.
It has, for example, been argued by Dennett (1991) and others that
"subjective experience"
as illusory and that once all
the brain processes associated with a given behavior are objectively described
there is nothing left to be explained.
Dennett's arguments, if correct, would eliminate the need to localize a
mind/brain interface or to solve a "binding problem." If, on the other hand, subjective
experience is a phenomenon in need of explanation, then the proposed
single-neuron theory, which shifts the locus of the mind/brain interface from a
macroscopic to a microscopic domain, might provide fresh avenues for addressing
Chalmers' "hard problem." In
the section of the paper entitled “Speculative Implications” I will provide a
specific example of how this might play out.
Other Single-Neuron Models:
Comparison of the single-neuron theory with other models that focus on single
neurons may help clarify just what the single-neuron theory does and does not
propose. For example, it is important
to distinguish the function of the single neurons in the single-neuron theory
from those referred to as "grandmother cells," single neurons whose
activation corresponds to the experience of a single memory (Thorpe 1998). The grandmother cell concept has in common
with the single-neuron theory the notion that a single neuron may contain
sufficient complexity to account for the complexity of a complete conscious
experience. It differs, however, in
that the grandmother cell is assumed, in the context of explaining memory
function, to be forever attached to a single experience, whereas the single
neurons of the single-neuron theory flexibly change their experiences over
time. Among other consequences of this
distinction, the limitations in processing capacity that plague grandmother
cell models, referred to variously as the "combinatorial problem"
(Singer and Gray 1995) or the "problem of exponential explosion"
(Gold 1999), do not apply to the single-neuron theory.
Zeki's Model: Another single-neuron based model was
described by William James (1890) in his discussion of what he called the
"theory of polyzoism or multiple monadism." According to this view, "every brain-cell has its own
individual consciousness, which no other cell knows anything about, all individual
consciousnesses being 'ejective' to each other." Along similar lines, Zeki and Bartels (1999) have recently
proposed that a primate visual brain consists of many separate functionally
specialized processing systems comprising hierarchical nodes that each generate
a "microconsciousness."
Neither of these schemes, however, explains how the individual
microconsciousnesses are able to induce macroscopically evident behavior. Zeki attempts to solve this problem by
suggesting that the many microconsciousnesses participate in a higher order
integrative activity that produces macroscopic consciousness. This, however, returns us to the received
view that consciousness is ultimately mediated by multiple neurons linked
through dynamic mechanisms. James, on
the other hand, considers the possibility that a single pontifical arch-cell by
itself might mediate the
totality of one's consciousness but he goes on to assert that there is no
evidence for such a single "center of gravity" in the brain. Both authors appear to overlook the
possibility that multiple separately and redundantly conscious neurons might
summate their outputs so as to produce the illusion of a single mind.
Bieberich's Model: Two other investigators have published
models featuring single-neuron consciousness.
Edwards (2005) proposes a model
of individual neuron consciousness not unlike the one presented here, but he
approaches the subject from the point of view of philosophy and physics and
offers no specific neuroanatomical mechanisms. Bieberich (2002), on the other hand, gives a detailed anatomical
account of his single-neuron theory. He
proposes a theory of "recurrent fractal neural networks" in which
spatial information encoded at the neuronal network level is reflected in the
spatial activity patterns exhibited in the dendritic trees at the single neuron
level. As with the single-neuron theory
proposed in the present paper, the individual neurons in Bieberich's model are
assumed to mediate the entirety of a subject's conscious experience at a given
moment.
Bieberich's theory has certain difficulties