It is a working assumption of cognitive science that all fast computations in the brain are done by neurons. However, this is only a working assumption. Life is built on marvellous and intricate mechanisms right down to molecular level, and some of these mechanisms (other than transmission of neural impulses) might be used for computation and information storage in the brain. We do not yet know enough about brain function to rule out the possibility. So we should from time to time question our working assumption, and explore the alternative — that the brain uses a hybrid of neural and non-neural computation.
However, there is no point in postulating non-neural mechanisms without sound biological reasons. Any non-neural mechanism must help the brain meet some strong selection pressure, in a way that neurons alone cannot.
Non-neural information storage might have particular advantages as a very short-term memory (working memory) and 3-D representation of the shapes and positions of objects in local space. Keeping track of these is vital for survival, giving rise to the strongest selection pressures on the brain; pressures so strong that they might drive the evolution of non-neural cognition.
I argue that purely neural mechanisms are not well suited to this time-critical task, and there are no candidate neural structures in the brain to do it. As an alternative, positions of objects might be represented by the wave-vectors of a wave-like excitation in the brain which couples to neurons for memory storage and retrieval. This has the potential to out-perform neural memory in information capacity, geometric accuracy and fast selective retrieval.
Is it implausible that neurons should couple to a wave-like excitation inside the brain? It is arguably no more implausible than their coupling to a wave-like excitation in the retina. Both evolved to meet the same massive selection pressure - the need to know where things are, as precisely as possible, at every moment. It would be biologically implausible if the precise optics and geometry of the eye were not matched by some precise geometric representation inside the brain; wave excitations, I propose, can do that job better than any neural mechanism.
We do not yet know what biophysical mechanisms might support these wave-like excitations in the brain. To look for them experimentally, we should have some candidate mechanisms in mind. One possibility is some form of high-temperature superconducting/superfluid state, such as the pumped polarisation excitations investigated by Fröhlich; there is evidence that these may occur in biological material.
Superconductors and superfluids are a Bose-condensed state; being frictionless, this state of matter can store information indefinitely at low energy levels, so is well suited to be a basis for memory. If present in the brain at low energy levels, such states might well not yet have been observed experimentally. It seems most likely that this working memory wave storage is located in the thalamus.
If, as proposed by Marshall (1989), phenomenal consciousness is a property of Bose condensates, and a Bose condensate is used for a 3-D working memory, we then have a very predictive and constrained theory of consciousness. It predicts both the limits of consciousness, and the form of consciousness within those limits. The predictions agree well with many properties of consciousness - for instance, that we are conscious of a spatial model of present local reality; the working memory, too, is a representation of present local objects in space.
Neural theories of the brain have two missing elements:
There is a remarkable coincidence between these two gaps - in that our conscious experience feels like a 3-D model of local space, which is just the critical missing functional element. This paper argues it is no coincidence, because the same ingredient fills both gaps - a non-neural storage mechanism in the thalamus, which which is the basis of both 3-D working memory and consciousness.
In order to arrive at a novel account of consciousness, we need to make several hypotheses. As some of these are unfamiliar, the reader may suffer from hypothesis overload, straining the credibility of the theory. However, at base the theory makes one hypothesis:
A. There is something non-neural going on in the brain; that something is a spatial working memory, and is the seat of consciousness.
This is the sine qua non of the theory, by which it stands or falls. The other more detailed hypotheses - wave storage of spatial information, a Bose condensate to support the waves, and the location of wave storage in the thalamus - are in some sense illustrative, ‘existence proofs’ that we can build a coherent theory along these lines in agreement with the data. While each detailed hypothesis is the current best candidate, any one of them might be replaced or modified without losing the essence of the theory. If they are replaced by better ones, so much the better.
Sections 1 and 2 argue on functional grounds (irrespective of consciousness) that a non-neural spatial memory may serve a vital biological function. Section 5 shows how this same non-neural memory can bridge the explanatory gap to consciousness, in good agreement with the facts. In between, sections 3 and 4 propose a possible mechanism and location for the non-neural component, introducing the detailed hypotheses above. The core of the theory is in sections 2 and 5, with the more speculative details in 3 and 4.
1. We should allow the possibility of non-neural information storage in the brain
1.1 Neurons are slow and have low output bandwidth; they may not be the best way to do all brain functions.
1.2 Neurons couple to signals at very low intensities, so they may couple to forms of energy in the brain which are hard to detect.
1.3 Information storage is simpler than computation; non-neural memory may be particularly feasible.
The neuron is a flexible general-purpose computing device, and connectionist models show how large arrays of neurons can perform a wide range of computations. We have come to regard the neuron as the general-purpose workhorse of the brain, for both computation and information storage.
However, neurons have limitations; they are slow devices, with a latency of around 10 milliseconds, and have a limited output bandwidth (of the order of 1 bit every 10 ms). Given this slow speed, it is hard to see how the most time-critical cognitive functions (typically involving vision and movement) are done in the very rapid times, of 100 ms or less, which we observe.
Maybe we will discover ingenious highly-parallel neural architectures which solve these problems. An alternative is to investigate those aspects of cognition where the problems are most severe, to see whether some hybrid neural/non-neural mechanism may do better than a purely neural design. That is the approach of this paper.
If such a hybrid mechanism exists, it is not clear that we would have detected it yet. Neurons can couple to other forms of information at very low energy levels (e.g at the one-quantum level, in the case of light). Since energy consumption in the brain is at a premium, we would expect any non-neural representation of information to use very low-energy physical effects, coupled to highly sensitive neurons. These low-energy effects might be hard to detect in living brains. In spite of recent advances, we do not yet know enough about the workings of the brain to rule out this possibility.
Where should we look for such effects ? I suggest we should look at:
2. The representation of local spatial information may require non-neural storage
2.1 Knowing the disposition of objects in local space is vital for survival, giving very strong selection pressure on the brain.
2.2 It requires a working memory for spatial relations, in an unrestricted 3-D metric representation of local space.
2.3 Having one master map of local reality is simpler and more efficient than just having multiple cooperating sensory maps.
2.4 For an unrestricted 3-D representation of local objects, neural storage has problems of information capacity and wiring.
To know what drives the design of brains, we must understand the strongest selection pressures; these probably centre on the control of physical movement. Within this function we can distinguish:
For (2), each movement must be appropriate to the spatial disposition of local objects and one's own body. For a small primate (such as our tree-dwelling ancestors) typical questions to be answered, in planning movements, are:
Q1. Can I jump onto that rock over there ?
Q2. If I move my paw from X to Y, will it hit anything on the way ?
Q3. Can he reach me ?
Q4. When the branch bends, were will I end up ?
To answer these questions, one needs to know the shapes, positions and movements of local objects around oneself. There has been massively strong selection pressure on our ancestors to meet this requirement - possibly the strongest selection pressure on the mammalian brain.
A small monkey moving through the forest canopy may cover his own body length and place a paw every 200 milliseconds. In this time, his view may change radically as new branches, obstacles and paw-holds arise. At each moment there are tens or hundreds of things in view — leaves and branches, parts of his body, peers, or other animals. He must not only recognise and classify these things (eg to avoid a rotten branch) but also know their shapes and positions, to answer questions like Q1-Q4.
Time is critical, and any error may be punished within a fraction of a second, at any second of the day. Even to place a paw wrong with probability 1 in 1000 means many, possibly fatal, falls per day - a huge deficit in fitness over a lifetime. It is lifetime fitness which determines the strength of selection pressures, and this depends on the number of challenges in a lifetime.
Reproductive behaviour may be tested once a year; food-gathering behaviour about once a day; social behaviour every few minutes; predator-avoiding behaviour every few hours. But the knowledge of objects in local space is tested every second. So the selection pressures on this facet of cognition are stronger than the pressures on any other, by perhaps two orders of magnitude. Even a 1% improvement in the speed or precision of these movement calculations would give a large increase in fitness.
The requirement for fast, reliable planning of movements is the strongest selection pressure on the primate brain, and so may be expected to produce the most remarkable designs.
Answering questions such as Q1-Q4 reliably (and thus making good choices of movement) requires a wide range of `what if' computations about the dynamics, positions and movements of local objects. These computations use the operations of three-dimensional vector geometry (for instance, to calculate the three-vector displacement between two points). Movement is geometry in time.
Input for these geometric computations comes from the senses, particularly vision. Sense data give geometric information about the positions of objects, which is used to plan movements. However, it is not enough just to have this geometric information as it arrives from sensory cortical areas; data from previous sense data over the last few seconds is also essential. As one's gaze shifts over a scene, information from one visual focus should not be immediately forgotten; it may prove useful a second later to plan some unforeseen movement.
For the best performance, therefore, we need a very short-term memory (a working memory) for the shapes, positions and movements of local objects.
This working memory might be stored in a variety of possible encodings, but the stringent demands of movement planning narrow the choices. Anything which loses geometric accuracy, or requires lengthy decoding, will thereby be less fit. I propose that the working memory is an unrestricted three-dimensional metric representation of objects in local space, which minimises both distortion and decoding time.
This representation needs to be:
Unrestricted: so it can model an unrestricted range of arbitrary objects and shapes - unlike a symbolic representation, which models only a restricted set of archetypal shapes denoted by its symbols. Real rocks and trees have arbitrary shapes, which must be modelled faithfully to plan movements.
Three Dimensional: because reality is 3-dimensional, and movements are executed in three dimensions; anything less than a 3-D representation would introduce dangerous ambiguities.
Metric: because movement questions, like Q1 - Q4, are metric questions, requiring actual distances and angles (or equivalent information) to answer them. This metric information needs to be retrieved directly and accurately from the representation, to answer the questions.
A Representation: because the number of movement questions which may arise, and the information required to answer them, is unlimited. A simple catalogue of facts would have insufficient capacity to answer all questions; we need a map-like representation, which implicitly contains all geometric facts.
From this it is clear that we require an analogue representation (Sloman 1974), or map of local space, rather than a symbolic representation. Symbolic representations, (particularly logic-based) have been much studied in the AI community (e.g Hayes 1985; Weld & de Kleer 1989), discovering their limitations for physical problems (e.g McDermott 1986). In this context, the two key limitations of a symbolic representation are (a) imprecision in representing arbitrary shapes and (b) time delays in reconstructing geometric information from a symbolic encoding. For instance, Marr's (1982) 3-D model is symbolic, and has these limitations.
The encoding of the visual field in the V1 visual cortex is an example of an unrestricted representation, modelling arbitrary shapes in two dimensions. Representations in later visual areas are typically more symbolic, and committed to specific purposes.
Given the massive selection pressure for fast, precise geometric computations to plan movements, it is clear that an unrestricted 3-D representation of local space (if feasible) would have big advantages and would be strongly selected for. It would not replace sensory topographic maps or symbolic representations, but would supplement them.
If it is used to plan bodily movements, the 3-D metric representation needs to integrate spatial information about the body (from somatosensory data) with information about other local objects (largely from visual data) into a consistent 3-D model of reality. Having to deal with at least two modalities of sense data, it might then be used to integrate all modalities - bringing together the texture and shape of a stone as felt by the fingers, its position as sensed by the arms, and its shape as seen by the eyes. Why this is useful can be illustrated by an analogy.
Your ship arrives at an unknown island, and you send out scouts to explore the resources of the island. How will you collate and cross-check the reports sent back by different scouts ? The best way is to draw a map of the island, and to reference every report onto this one map. That is the only good way to tell if two river sightings are of the same river, or different rivers. An unrestricted map can accommodate every kind of report, and is the best basis for comparing all types of information; a symbolic catalogue of reports could not do the same job, nor could a set of separate maps made by each scout. The geometry of your map must match the geometry of the island; if it happened to be a 3- dimensional island, you would need a 3-dimensional map.
This analogy shows the merits of having one `master map' of reality in the brain where all sense data can be collated, and conflicting interpretations resolved. Neuroanatomists have looked for such a master map in the brain and not found it. There are many two-dimensional topographic maps in the cortex, but none of them seems to be a primus inter pares, or convergence region for all the others (see e.g. Damasio 1989, Crick 1995); because we have not yet found any master map, the prevailing view is that our model of reality is built up cooperatively by the many cortical maps.
However, there are good reasons why one `master map' is more effective than equal cooperating maps; the single-map desert island story has sound design and performance reasons behind it. To see this, consider the two alternative neural architectures, shown in figure 1 below.
Figure 1: Connectivity required for (a) master map and satellites, and (b) cooperation of many maps
The two diagrams each show six cortical topographic sensory maps, (a) with a single central master map, and (b) without one [1]. The advantages of the `master map' architecture are:
(1) Translation Complexity : To detect a mismatch between any two of the maps in figure 1(b), feature descriptions must be translated from one map to the other, and any feature in one map must be correctly aligned with its image in the other. As each map is a 2-D topographic projection of 3-D reality, the translations between them are complex. Interfaces between the sensory maps are not just simple neural excitation paths, but need to convey a complex `mapping between maps' faithfully. There are N(N-1) such mappings (30 in this case).
In the master map architecture of figure 1(a), with N sensory maps there are only 2N translations (12 here); and each translation, being a translation from a 3-D model of reality to a 2-D sensory field, is quite simple - simpler than the translation between two sensory fields.
Having N(N-1) complex translations rather than 2N simpler translations is a major increase in the design complexity of the brain. You might think that after so much evolution, brains are complex, so we can afford the complexity if it performs better. However, extra complexity is very costly when coping with changes, either within a lifetime or over evolutionary time.
We know that cortical sensory maps can change within a lifetime. If so, the mapping between maps must change with them. To change one sensory map, and to change one translation each way to the central master map, seems like a feasible learning process. Without a master map, even to change one sensory map requires changing all its complex translations to the other sensory maps in step, which is much harder and slower to do. Design simplicity has tangible fitness benefits.
(2) Required calculations : The task of perceptual processing is to find the most likely state of the world, for given sense data - or to find that state of the world which maximises the likelihood, P(state|sense data). Through Bayes' theorem, this is a product of terms P(sense data|state) for the different modalities of sense data, weighted by prior probabilities of states. In this context, the `state' is the 3-D model of local space in the centre of figure (1a), so the radial connections are just the links required to calculate each P(sense data|state). While the links of figure (1b) might be used to compute P(sense data|sense data), that is just not what brains need to calculate.
It is known that many aspects of cognition have a near-Bayesian form, including some aspects of early visual processing (Barlow 1978,1980; Burgess 1990) and high-level functions such as making foraging decisions (Stephens & Krebs 1986); it has also been shown that in general, near-Bayesian cognition gives greatest fitness (Worden 1995). The radial topology of figure (1a) is well suited to do this; the many-maps topology of figure (1b) is not.
(3) Fast Decision-making : When a committee has a chairman, he can decide any issue by a straight vote; but with no chairman, the decision is reached by a complex process in which local coalitions form, and then vote against other coalitions. Similarly, in the distributed decision-making of figure 1(b), local groupings of sensory maps may agree with one another about some feature which is later shown to be wrong - overridden by the knowledge from some other maps. One sensory map may have several successive `changes of mind' during this settling process.
So even for a simple binary decision, the `autocratic' master map solution has a speed advantage of the order of log(N) over the cooperative solution. This speed advantage is all the greater for complex interdependent sets of decisions, needed to construct a consistent model of reality from sense data.
Therefore the master map architecture has advantages of efficiency and simplicity over the multi-map alternative. If there is such a master map, which is an unrestricted 3-D representation of local space, how might it be realised in the brain ?
There are few published accounts of how an unrestricted neural 3-D neural representation might work, and equally few candidate neural structures in the brain. This lack of candidates may reflect two problems:
Information capacity: The most popular neural models of very short-term memory use an encoding in the firing rates of neurons, maintained by positive feedback - rather than short-term synapse modification. This encoding gives an information capacity of the order of 1 bit per neuron (versus 1 bit per synapse, which is possible for storage by synapse modification).
For 2-dimensional analogue representations, such as the visual field encoding in V1 cortex, a cost of 1 neuron/bit is acceptable; to represent the visual field with a spatial resolution of 1 part in 10**3 requires of order 10**6 bits, or 10**6 neurons. However, for a three-dimensional analogue representation with resolution one part in 10**3 , where each neuron represents some small region in space, the cost would be of order 10**9 neurons, which for a small mammal is prohibitive [2]. Most of this storage capacity would be wasted at any instant, because most of local space is empty.
To avoid such a high cost of storage, neural models of memory based on synapse modification are necessary.
Wiring Topology: The cortex is a 2-dimensional structure, so it can easily store 2-D unrestricted analogue representations. The third spatial dimension of the brain is then used to take information into and out of the representation. To represent 3-D information in the same analogue fashion - where the physical position of a neuron corresponds to the physical position of the thing it represents - would require a three-dimensional clump of neurons, containing of order 10**9 neurons. There would then be severe problems of wiring, to get the information into and out of the central neurons, through the barrier of outer neurons.
Also, it is hard to design a three-dimensional clump of neurons without local inhomogeneities (such as blood vessels) which would make gaps in the representation.
So the direct analogue approach (of representing local space by a neuron-filled space in the brain) appears not to work, and we are forced to consider alternatives.
We may try a more distributed connectionist representation, where any one neuron does not just represent one point in space. This loses two benefits of the direct analogue representation:
Another alternative is to supplement a 2-dimensional unrestricted representation (as in the V1 visual cortex) with depth information for every point. This would be something like Marr's (1982) 2½-D sketch. It is economical in storage and could encode objects of any shape. Its drawback is that complex decoding is required to recover basic three-vector information (e.g the 3-vector displacement between two points) for motion calculations; this introduces extra delays and errors.
I do not imply that a neural/connectionist encoding of unrestricted 3-D spatial information is impossible - just that it presents serious design and performance problems which have not, as far as I know, yet been addressed. It is a challenge to connectionist researchers to find such mechanisms. Meanwhile it is worth investigating the alternative, that a non-neural mechanism is used in the brain for this purpose.
3 A wave excitation in the brain may store local 3-D spatial information
3.1 Wave excitations can store large amounts of 3-D information efficiently and faithfully
3.2 Spatially selective retrieval can deliver the information needed for the geometric computations of movement
3.3 3-dimensional wave storage may occur in the thalamus
A wave in a three-dimensional medium may travel in any direction, and has a variable wavelength. In this respect it has three degrees of freedom, as does a point in three-dimensional space. Therefore a wave may represent a point. The direction of the wave represents the direction of displacement of the point from some origin, and the inverse wavelength represents the distance of the point from the origin. This is the basis of the unrestricted representation of points in space which I propose is used in the brain.
Suppose there is a medium in some region of the brain which is capable of supporting wave-like excitations. This medium need not be separate from the neurons and glial cells which make up much of the grey matter, but might consist, for instance, of molecules distributed in these cells. Some physical parameter of the medium (such as its electric polarisation), denoted by E, can have a wave-like dependence on position x in the medium, E = E0cos(k.x). Here k and x are both three-dimensional vectors, and k.x is their scalar product. k is the wave vector of the excitation, and its magnitude is inversely related to the wavelength l by k=2pi/l.
If the wave equation in the medium is linear, there can be very many different wave excitations in the same region, with different wave vectors k, without any interference between them. Therefore the waves in the medium can serve as a representation of spatial information. A small object at a position r relative to the animal is denoted by a wave in the medium with wave vector k = µr, where µ is a constant. Objects at different positions are represented independently by excitations with different wave vectors in the same region. The wave excitation is the Fourier transform of the distribution of objects in space. It can represent an arbitrary set of points in three-dimensional space; or equivalently, objects of arbitrary shape.
To represent points at very large distances would require very large wave vectors (i.e very small wavelengths) which would not be feasible. Therefore the wave storage may use a non-linear radial distortion; for instance, the stored distance deff might be related to the actual distance d by deff = d0 tanh(d/d0). This means that for distances much smaller than d0 , the representation is essentially undistorted with deff = d ; but for larger distances there is a progressive radial compression, so that the stored distance deff is never larger than d0 .
This wave-vector representation of position information is similar to that used in a hologram, and has similar advantages:
Neurons must couple to the wave medium, to store and retrieve information. While we do not yet know the biophysical mechanisms, there are useful things to say about the form of the coupling.
Suppose that certain neurons have transducers in the medium (analogous to the light-receptive structures in rods and cones) which couple to the oscillations.
To retrieve information about an object at some spatial position r, we need to observe the component of the oscillation in the medium with wave vector k=µr - taking an average of the excitation E over many points, weighting each point x by the factor cos(k.x). The more points x included in the average, the greater the precision with which the the object's position r is defined. To get large numbers of points, it seems likely that the transducers are associated with synaptic structures, rather than with cell nuclei.
Wave retrieval has a wiring advantage over the direct `representing positions by positions' neural storage discussed in section 2. To retrieve information about objects at all points in the represented space, we do not need transducers at all points in the medium. A sampling of points is sufficient, and it may be made compatible with any wiring constraints.
Suppose the medium is permeated by neurons, each one having transducers distributed on its dendrites so as to be sensitive to a particular wave vector k= µr , and thus to store or retrieve information about any object at position r. The drawback of this is that to cover 10**9 possible points r , we need 10**9 neurons - just as many as a pure-neural solution would have used for storage.
There are two possible solutions:
(a) Tunable antennae : Each neuron is an extended antenna, summing contributions from transducers on its dendrites to find the excitation with a given wave vector. The sensitivities of different tranducers could be altered dynamically (e.g by synapses from other neurons) to tune or `steer' the antenna to different wave vectors. This can be used for both storage and retrieval - transmission and reception.
(b) Non-linear `searchlight' retrieval: The neural tranducers are sensitive to the wave excitation E in a non-linear manner; for instance, if their response is sensitive to the local energy density E**2 . Then to retrieve information about objects in the vicinity of some spatial position r0 , the medium may be stimulated by an extra `searchlight' excitation with wave vector k0 = µ r0 . Existing memory traces for spatial positions near r0 (i.e. excitations with wave vectors near k) will beat with the extra excitation to produce slowly-varying changes in E**2. Neurons with receptors tuned to observe this slowly-varying component of E**2 will observe the memory traces for objects in the neighborhood of r0 .
Either (a) or (b) removes the need for very large numbers of retrieval neurons. The non-linear method, while selectively retrieving information about objects in a small spatial neighborhood, also computes the three-vector displacements of points relative to the centre of the neighborhood. This spatially invariant retrieval is of great help for many important computations.
The laws which determine the movements of objects are spatially invariant. The same laws apply to an object 20 feet away as to an object much closer, so the same computations are needed in both cases to plan movements. These computations are much easier if their input data are presented to them in a spatially invariant manner - if similar groups of objects give similar retrieved memory traces, wherever they are. It has all the advantages for navigation of a faithful map over a highly distorted map. That is just what the nonlinear retrieval mechanism gives.
It is useful not only for dynamic-geometric calculations of movements and their effects, but also for spatially-invariant object recognition. Animals need to recognise and classify objects independent of their position, and a retrieval method which automatically subtracts out the position r0 of the centre of an object, giving information about the positions of its features (rn - r0) relative to the centre, gives a head start in doing this. Neural nets, while being good for object recognition, are generally not good at spatially-invariant object recognition without some assistance such as this provides.
The wave storage mechanism meets all the requirements identified for a spatial working memory. It is:
The main inputs and outputs of the 3-D representation are summarised in figure 2.
Figure 2: Information flows to and from the unrestricted 3-D spatial representation
Each box in this diagram denotes both information storage and some associated processing operations. The correspondence with figure 1(a) is as follows: the master map in the centre of figure 1(a) is the unrestricted 3-D spatial representation,and the satellite maps of figure 1(a) are all in the `object recognition and representation' box of figure 3.
If there is an unrestricted 3-D representation of local space, where is it in the brain? The mechanism requires a three-dimensional region within which wave excitations can occur. This region should have roughly equal size in all three dimensions, to support wave vectors defined with good precision in all directions.
The cortex, being a two-dimensional structure, is not suitable. White matter (myelinated axons), while three-dimensional, is also unsuitable because even if it could support the excitations, the required two-way information coupling to neurons seems unlikely. This narrows the search [3] down to the grey-matter nuclei such as the thalamus and brainstem structures. Of these, the thalamus seems a good candidate because:
In this model, the wave excitation may permeate the whole volume of each thalamus (or a large part of it). Neurons in specific nuclei of the lateral thalamus act as transducers conveying sense data into the wave.
Three lines of evidence support the idea that the thalamus is the site of the wave storage:
1. Attention-Focusing Mechanisms: If the 3-D representation has a selective `searchlight' retrieval mechanism as described above, there need to be control mechanisms to direct the searchlight. There is mounting evidence (reviewed by Newman 1995, 1996) that the thalamic reticular nucleus, which surrounds the thalamus and acts as a gateway, serves just this purpose - directed by inputs from three brain regions which are expected to control attention - posterior cortex (when salient objects are recognised), the midbrain reticular formation for orienting to novel stimuli, and prefrontal cortex for social/emotional control of attention (Scheibel 1980).
2. Evolutionary Timescales: The selection pressure to understand the shapes and places of objects in local space is very strong. But it has not just become strong in recent evolutionary time; it has probably been a very strong pressure for almost half a billion years. We would not expect the brain's response to this selection pressure to be localised in recent brain structures, such as the neocortex. We would expect the 3-D representation of local space to be located in one of the older structures, and in this respect the thalamus seems suitable - being an older, mid-brain structure yet having copious and direct access to sense data.
3. Distinctive Synaptic Structure: Wave transducers are probably associated with synaptic structure; there are distinctive large synaptic aggregations in all thalamic nuclei, involving the dendrites of relay cells and interneurons (Jones 1985; Steriade et al 1990).
We can even sketch a possible thalamic architecture for wave storage and retrieval. Each thalamic relay cell receives dense subcortical input on its proximal dendrites. Its distal dendrites act as an antenna, to store and retrieve from the wave excitation. This antenna is dynamically tuned to specific wave vectors by corticothalamic afferents (which terminate on the distal dendrites) and interneurons. Thus the relay neurons both pass information on to the cortex, and store it in the wave excitation; cortical feedback tells them where to store it [4] (i.e where in represented space to put each sensory input).
The same relay neurons may also retrieve from the wave storage, using the same corticothalamic tuning signal; possibly the 40 Hz oscillations control a store/retrieve cycle. Non-specific nuclei may have a retrieval-only function such as the non-linear `searchlight' retrieval.
The thalamus is not currently favoured as a site for a master map of local space. One reason for this is that the specific thalamic nuclei (which have reciprocal connections to many cortical regions) have few interconnections with one another, and so do not seem to be part of any integrated map (Mesulam 1985). However, in this theory all the specific thalamic nuclei couple to the wave storage which permeates the thalamus, and which is the master map.
Because the thalamus has not been thought to be the site of a master map, it is currently seen as a coordinator of many representations distributed in the cortex. Supporting this, in white matter the corticocortical connections outnumber thalamocortical connections by about 10:1 (Newman 1996). However, it is not clear how many of these corticocortical connections are intra-map rather than inter-map (e.g between successive stages of visual processing), or are connections to other non-sensory regions such as prefrontal or entorhinal cortex. Thalamocortical connections are still profuse enough to register sensory topographic maps onto a master map.
The searchlight selective retrieval mechanism is reminiscent of Crick's (1984) searchlight hypothesis. The hybrid cognition theory implies a two-way searchlight, pointing both inwards (to parts of the wave representation in thalamus) and outward (to related parts of cortical representations) in a coordinated way.
In this connection the neural net model of Taylor and Alavi (1993) is particularly interesting. This has a winner-take-all competitive network in the thalamic reticular nucleus, where the winner is a periodic pattern with some spatial wavelength over the whole reticular nucleus. A spatial wavelength corresponds to a specific position in represented space. So Taylor & Alavi's model is a candidate for the searchlight mechanism in this theory.
Having one master map of reality in the thalamus radically alters the binding problem, of relating neural activity in different cortical areas. Because neural actvity in different sensory maps can be bound by their links to the common thalamic map, there is no need, for instance, to assume that the `40 Hz' thalamocortical oscillations have this role (Crick & Koch 1990; Llinas et al 1994; Hardcastle 1994). They may then have a simpler role in synchronising iterative refinements of the 3-D model between the thalamus and the cortex - acting as a timebase for these computations - or in controlling a store/retrieve cycle for the wave memory.
4. Bose-condensed states may support the 3-D wave storage
4.1 A superconducting/superfluid state, or Bose condensate, insulates information from thermal fluctuations, storing it with little energy input.
4.2 Fröhlich-type polarisation oscillations are one candidate high-temperature superfluid state, and may occur in biological material.
4.3 They may be used as a memory mechanism, at low intensities which would be hard to detect in the brain
High-temperature superconductors show us how a remarkable state of matter can exist near room temperature in complex materials, for reasons we did not suspect until it was observed.
This remarkable state - the Bose-Einstein condensate, where very many quanta are in the same quantum state - is frictionless, so it can store information for long periods, insulated from thermal noise (Tilley & Tilley 1990). This makes it a candidate basis for biological information storage.
Could there be some form of high-temperature superconducting or superfluid state in the brain, which acts as substrate for the wave storage described above? There are three reasons why we should not rule out this possibility too hastily:
There is already one proposal for a superfluid-like mechanism in biological material. This is the coherent polarisation oscillation investigated by Frohlich (1968). While his theoretical analysis is controversial, and some other, (perhaps more complex) Bose condensed state might equally be the basis of a `superconducting' memory in the brain, it is worth outlining the Frohlich mechanism, as one example of how Bose condensed states might exist in the brain.
Fröhlich noted that biological material contains many highly polarisable large molecules, particularly in cell membranes, so that these molecules form a polarisable medium. There may then be longitudinal oscillations of polarisation in the medium, which propagate by the influence of the longitudinal electric field on the molecules.
These oscillations would typically have very high frequency (of the order of 10**11 Hz) and could have a wide range of wave vectors; their frequency depends only weakly on their wavelength.
Fröhlich suggested that energy sources in cells might pump energy into these oscillations, which could be used for several purposes - for instance in regulating cell division. Since his suggestion, experiments have shown that biological materials (e.g yeast cells) are strongly affected by radiation at the relevant frequencies; much more strongly than expected from purely thermal effects, and often with a very sharp frequency dependence (Grundler et al 1982).
Fröhlich investigated theoretically of the physics of the polarisation oscillations induced by biological pumping. The oscillations are Bose quanta, and obey Bose-Einstein statistics; therefore when any quantum state is highly populated, the tendency of more quanta to move into that state is increased. He found that if the rate of energy input into the ground state exceeds a certain threshold, then a finite fraction of all the quanta go into the ground state, giving a form of Bose-Einstein condensation.
The Bose-Einstein condensation is a unique and highly distinctive state of matter with very many quanta in the same ground state. It occurs in superconductors and superfluids, where (because thermal fluctuations cannot knock one quantum out of the ground state) the state is frictionless. Superfluids have no viscosity, and superconductors have no resistance. Therefore they can store large amounts of information (e.g in quantum vortices) indefinitely. Any such Bose-condensed state - including Frohlich's polarisation waves, if they occur in the brain - is therefore well-suited for the very short-term memory function.
In this model, polarised molecules in the brain (e.g. in cell membranes) act as a medium supporting polarisation oscillations. Some form of energy pumping (from neurons or glial cells) creates a Bose-Einstein condensate of these polarisation quanta; or there may be some more complex, unknown Bose condensation mechanism. There are neural transducers which modulate the form of the condensate (much as the container walls alter the shape of superfluid helium) in wave-like patterns which then persist, and other neural transducers detect this shape.
The whole mechanism would act like a kind of inner eye in the brain, in which neurons both create and observe a highly coherent and persistent inner light. This light acts as a memory and representation of local space.
This sketch clearly leaves huge gaps and unanswered questions. It is offered to illustrate that wave-like storage in the brain may be feasible, that candidate biophysical mechanisms can be found, and that they may be worth investigating both theoretically and experimentally. Serious biophysical investigation would be needed to find convincing candidate mechanisms - possibly using Bose condensates completely unlike Frohlich's.
5. Bose-Condensed states may be the basis of consciousness
5.1 In a hybrid theory of cognition, it is possible that only the non-neural component is conscious.
5.2 If any Bose condensate consciously experiences its own quantum state, then for the condensate in the brain, that experience is a 3-D model of local space.
5.3 This makes clear predictions for the limits and the nature of conscious experience. They agree well with the evidence.
Neural theories of consciousness give us no clear grounds to understand how some specific pattern of neural firing gives rise to phenomenal consciousness - our experience of what it feels like to be us - while most neural activity does not.
However, if there are two distinct computational mechanisms in the brain (neural and wave storage) then we have a chance to understand our phenomenal consciousness in this two-component theory - by assuming that only the non-neural component is conscious. How well does this two-component theory fit the facts ?
To make the theory specific, we assume, following Marshall (1989) a law of nature that (1a) Bose-Einstein condensates are conscious. This is a plausible candidate for a law of nature, being simple in form, and relating consciousness to a very fundamental and pure state of matter. Consciousness arises from very many quanta in an identical quantum state.
We further assume that (1b) the form of consciousness experienced is the form of the shared quantum state. From this we expect the Bose-condensed excitation which is the master map in the thalamus to experience its own form - a map of local space.
Because we have tied our account of consciousness to a simple law of nature, and defined the form of consciousness by (1b), we now have very little room for manoeuvre in explaining its properties. They have to be the properties of the Bose condensate, and of the unrestricted 3-D spatial representation; in describing what phenomenal consciousness is like, the buck stops here. There is no homunculus to observe the condensate, and to somehow make its properties right.
If our theory gets those properties wrong, there is little we can do about it; the theory is easily falsifiable. But, as is shown by the comparisons below, the theory seems to get them right.
In the tests which follow, recall that we are concerned with phenomenal consciousness, the `what it is like' aspect, rather than, for instance, access consciousness (having information available in the brain for other uses, such as verbal report) (Block 1995). Access consciousness can be measured in animals and in third-person psychological experiments; phenomenal consciousness is first-person experience. The data for most of the comparisons below is your own experience.
(1) Consciousness is about sense data of all modalities: The basic character of phenomenal consciousness is an awareness of sense data - sights, sounds, and feelings in the body [5].
The 3-D spatial representation acts as a master map for the integration of all sense data. So we expect all sense data to figure in consciousness.
(2) Consciousness is about the present moment: The bulk of conscious awareness has a time-window of just a fraction of a second, referred to as `the present'.
This is exactly what we would expect of a very short-term memory, whose purpose is to keep the information from sense data (and its processing) available for use in the next few moments. The timescales match closely.
While we can populate consciousness with memories, plans and imaginings, these are only a pale shadow of present-moment awareness. This, too, is expected, because memories are reconstructed from stored traces with much less information content than live sense data.
(3) Consciousness is located in space : Visual awareness is quite precisely located in our internal conscious space; awareness of sounds and bodily sensations is sometimes fairly precisely located, sometimes more blurred.
We would expect each form of sense data to be as precisely located in the 3-D representation as necessary for sensory integration and movement calculation - visual data very precisely, sounds rather less so and bodily sensations even less so. This is just what we feel.
(4) Consciousness is an undistorted spatial model of reality: We experience not just disembodied qualia, but located qualia with shape - defining the shapes and places of real rocks, trees, or bodies, vivid and undistorted. This may be summed up in the formula: “Consciousness = shape * qualia” , where the beautiful and precise `shape' part of this formula stands in need of explanation, just as much as the `qualia' part.
Consciousness might not have been so transparent and precise. It might have been a set of symbolic labels for `chair', `dog' etc. distributed in space; or it might have been a distorted map, so that the experienced shape of a tree changes as you approach it. It is neither of these; phenomenal consciousness is an undistorted metric map of reality now.
A 3-D geometric model used for planning movements needs to be undistorted in just the same way, to make movement planning `plane sailing'; so we expect the consciousness within it to be undistorted. That is perhaps the most specific and well-confirmed consequence of this theory.
(5) Phenomenal consciousness correlates with access consciousness: Empirically, being conscious of something correlates with that information being available in the brain for many kinds of functional processing (eg for making verbal reports, or for memorising) which do not necessarily have anything to do with phenomenal consciousness (Block 1995).
The unrestricted 3-D spatial representation (which by hypothesis, is the seat of phenomenal consciousness) makes information available to many other subsystems in the brain (leading to access consciousness); so the theory predicts that the two will be correlated.
(6) Consciousness has a variable focus of attention: Within our general awareness there is a changing focus on particular spatial sub-regions - which may be a part of the visual field, or a part of the body. Consciousness is more vivid and clearly defined within this focus.
The `searchlight' selective retrieval of section 3 implies that regions of the 3-D model have more intensive retrieval to cortical areas (producing a focus of access consciousness), accompanied by more intensive feedback from those cortical areas into the 3-D representation - sharpening the model in those regions and producing a focus of phenomenal consciousness. So we can understand how the foci of access consiousness and phenomenal consciousness move in step.
(7) There is a sense of unity in consciousness: We feel that the continous stream of overlapping qualia are all happening on one conscious `stage', not on separate screens of a multi-screen cinema. This matches with the Bose-Einstein condensate, which is a single quantum state, shaped to reflect the model of the local world [6].
A key role of the Bose condensed working memory is to create one unified model of the world from all the modalities of sense data; so we expect that the experience within the Bose condensate is experience of a unified reality.
(8) Much brain activity is not conscious : For instance, all the processes which go to generate or understand language, and many gradual learning processes, are completely outside consciousness. In this theory, those are all purely neural processes, so are not expected to be conscious.
(9) Damage to the thalamus interrupts consciousness : Clinical evidence shows that the thalamus, in particular the Intra-Laminar Nuclei, is essential for waking consciousness; no region of the cortex, and very few other brain areas, have this property (Bogen 1995). This clearly agrees with a theory in which conscious awareness is located in the thalamus.
(10) Diversity of Qualia: In the simplest Bose-condensate theory, one might expect a monochrome form of consciousness; one quale distributed with varying intensity over space. However, as the unrestricted 3-D spatial representation gathers together data of many sense modalities, there would be strong functional advantages if it were able to distinguish between memory traces of these different modalities.
Depending on the nature of the Bose condensate, this might be done by having several different kinds of excitation in the same Bose condensate [7]; whether or not this is possible will depend on properly worked-out biophysics of the condensate. Then different excitations, with the same wave vector, would represent different types of sense data (different qualia) at the same represented position.
(11) Quality of Qualia: If the account of diverse qualia in diverse excitations, sketched above, were viable, we might then extend law (1b) to have the form `excitation 1 leads to quale 1; excitation 2 leads to quale 2...' and so on, where `quale 1' might stand for `my sensation of green', etc. Law (1b) would then contain a set of qualia-valued constants of nature, each one linked to a type of excitation of the Bose condensate. However, qualia-valued constants are not communicable or shareable, so this extension of the law would not by very useful or publicly testable.
In properties (1) - (9), the theory is in remarkable and unforced agreement with the main facts of consciousness. The 3-D spatial working memory is about just the same things that consciousness is about - no more, and no less - and has a very similar character.
In the pre-war study of nuclear beta decay, something was missing. Some of the mass-energy of the initial state could not be found in the decay products. Taking this seriously led to the discovery of the neutrino, a particle with interactions so weak that it can travel through the earth without hitting anything; a particle which would never have been discovered directly at the time.
I have argued that for neural theories of cognition, something equally important is missing:
This is a bit like beta decay without the neutrino, or Hamlet without the prince. We can either press on hopefully with the neural theory, or take the gaps seriously as clues to something missing - something which, like the neutrino, may be so subtle as to have evaded detection.
I propose that one missing ingredient fills all these gaps - a wave-like representation of local space, stored in the thalamus, using some superfluid-like state of matter. This provides the metric map for motion calculations and the master map for integration of sense data. If we then assume that consciousness resides in Bose-condensed states, this gives a highly constrained and predictive theory of consciousness. It seems to agree well with the evidence - it successfully predicts what is conscious and what is not, it predicts the unity of consciousness, and predicts much of its phenomenal form.
It is reasonable to suppose that the wave excitations in the thalamus use low energy levels and couple to neurons in subtle, unobtrusive ways - so that, like the neutrino, they are hard to detect directly. However, the weak point of this theory is still the fact that we have not directly observed any such wave excitation in the brain.
To confirm or refute hybrid cognition, therefore, the priority is clear. We need either to observe the missing element directly, or to prove that it does not exist in the brain. To do either requires detailed biophysical and neuroanatomical analysis of possible mechanisms. The physical properties of candidate mechanisms - their wave modes, frequencies, and intensities - will determine the experimental means of looking for them, and will determine whether they are compatible with the known structure of the brain. A key challenge is to find the transducers whereby neurons in the thalamus couple to the excitation.
Such biophysical and neuroanatomical analysis might narrow the area of search quite rapidly. But as the mechanism might be as complex as high-temperature superconductivity, muscle contraction or DNA replication, some subtle and deep theoretical analysis may be required. The biophysics of hybrid cognition may be just as profound and beautiful as the biophysics of life, and no easier to unravel. If this proves to be the basis of consciousness, the effort will be worthwhile.
This theory is not so much as a radical departure from modern neuroscience, as (potentially) a radical addition to it. If true, it will enhance neural theories of cognition and consciousness, with a powerful extra ingredient - increasing, not supplanting, their explanatory power.
For instance, recent neural theories of consciousness can be regarded, from this viewpoint, as theories of how neurons shape conscious experience, even if neurons do not directly create it. The new ingredient of wave storage may require a limited reinterpretation of those theories, but many of their key insights will still stand; they were in fact the origin of several features of this theory. Space permits only the briefest mention of some major influences.
In a series of seminal publications, Crick and Koch (eg Crick 1984; Crick & Koch 1990; Crick 1994; Koch & Crick 1994) have developed neural theories of visual awareness which, like this theory, stress the importance of working memory and the thalamus. This theory may provide the link to phenomenal consciousness which, as Crick and Koch agree, is currently not defined for their model.
The Global Workspace (GW) theory of Baars & Newman (Baars 1988; Newman & Baars 1993) also shares several features with this theory. Their workspace acts as a kind of `blackboard' for broadcast of the results of computations between diverse neural expert modules, and thus has a role similar to the master map of section 2; although they see the workspace as distributed in cortex, coordinated by the thalamus. In their theory, as here, the global workspace is linked to consciousness.
Damasio's (1994) theory is important in stressing the importance of somatosensory data, and a `body model', in both consciousness and cognition. In this theory, a body model is part of the 3-D model of local space, and so a part of consciousness.
Gray's (1995) theory of consciousness and Taylor's (1991) relational model both stress the importance of stored memories for conscious experience. In this theory, stored memories are essential for object recognition in the cortex, which in turn helps to define the 3-D model in the thalamus - and so shape the form of consciousness. But long-term memory per se is not central to the existence of consciousness experience.
In all these theories, neural activity tells us important things about the nature of consciousness; but the key link to consciousness itself is still to be filled in by some new law of nature. This theory proposes a possible link, through an ultra-pure quantum state.
Many authors since Wigner (1961) have suggested that consciousness is linked with quantum effects. The link to the Bose-Einstein condensate was proposed by Marshall (1989), and has been developed by others (Lockwood 1989; Penrose 1994; Hameroff 1994; Cairns-Smith 1995). This theory is distinctive in that (a) the quantum state serves a vital biological purpose, (b) the form of the quantum state directly and simply defines the form of consciousness, and (c) it does not yet depend on any disputed interpretation of quantum mechanics.
The philosophical viewpoint of consciousness in this paper has been strongly influenced by the work of Chalmers (1995,1996); in particular, by his clear analysis of the nature of phenomenal consciousness, his insistence that there is something there to be explained, which cannot be argued away; and the need for some new and simple `bridging' law of nature to link consciousness to physical phenomena. However, this theory diverges from the later parts of Chalmers' work, in that the proposed simple law of nature depends on a fundamental quantum state, not on any `informational' configuration of neurons as he conjectures.
If consciousness resides in a Bose condensate, which by definition is an indivisible single quantum state, then it negates many philosophical thought experiments, including the silicon substitution `dancing qualia' experiments discussed by Chalmers; a Bose condensate cannot be incrementally replaced.
Aspects of this theory are unfashionable. Particularly so is its localisation of consciousness within one region of the brain; a recurrent theme in a recent electronic colloquium was, as summarised by Newman (1996) quoting Grace (1996), that “a localisation of such higher processes as consciousness and binding to a single brain region or even a single brain system is not likely to be the most fruitful approach” .
In a neural theory, this view is understandable, perhaps unavoidable. Whenever you look at a small enough region or group of neurons, you find nothing which seems capable of creating consciousness, or seems much different from other (unconscious) groups of neurons. However, distributing the problem across the whole brain has not solved it; we may still need to consider radically different, possibly localised solutions.
In proposing one master map of reality, this theory diverges from a mainstream view in neuroscience, that the brain's model of reality is built up by co-operation of many topographic maps in the cortex. The many-maps architecture has drawbacks of performance and complexity, but has become the mainstream proposal just because, in a pure-neural theory, there is no place in the neuroanatomy of the brain for a master map. Hybrid cognition allows such a place, in the thalamus.
A many-maps theory of cognition leads to a multiple-drafts theory of consciousness, as in the work of Dennett (1991) or Damasio (1994). Since the multiple drafts are in many 2-D topographic and symbolic representations, it is entirely unclear how the resulting conscious experience has its undistorted 3-D character.
It is as if four people, conversing over telephones in English, Finnish, Japanese, and Serbo-Croat, should somehow cause conscious experience in Sanskrit. One may ask - why Sanskrit, and not Swahili? Equally, why does conscious experience have 3-D Euclidean geometry, not 4-D Riemannian geometry ? Experience without a closely matching physical representation seems highly implausible and ungrounded; it asks too much of the required bridging law between physical phenomena and conscious experience.
Dennett's key argument against a single-draft, `Cartesian theatre' model of consciousness is the argument of the homunculus: if the Cartesian theatre requires an observer, or homunculus, then where is the consciousness in that homunculus ? Any answer seems to lead to infinite regress.
In this theory, consciousness resides in a coherent oscillation, which is a kind of inner light in the brain. By assumption, this light contains consciousness, and needs no observer. Neurons in the thalamus (the `inner retina') create and observe the inner light. We are conscious not because our neurons, or any homunculus, observe the light - but because we are the light.
If that account turns out to fit the facts, then it is no more mystical than the inverse square law of gravitation. In accounting for consciousness, as for gravitational motion, the buck has to stop somewhere; and making it stop in the wave storage seems to agree well with the facts of consciousness.
However, the key test of this theory will be to observe the wave storage mechanism directly in the brain. If it were found, I think the case for linking it to consciousness would be very strong. Even if wave storage is not found, and this theory turns out to be completely wrong, it may still stand as an example of what a simple, predictive, testable theory of consciousness would look like. That in itself is perhaps useful.
Baars, B. J. (1988) A cognitive theory of consciousness, Cambridge University Press
Barlow, H. B. (1978) The Efficiency of Detecting Changes in Random Dot Patterns, Vision Research, 18, 637-50
Barlow, H. B. (1980) The Efficiency of Perceptual Decisions, Philosophical Transactions of the Royal Society (London), B290, 71-82
Block, N. (1995) On a confusion about a function of consciousness, Behavioral and Brain Sciences 18:227-47
Bogen, J. E. (1995) On the Neurophysiology of Consciousness. Consciousness and Cognition, 4:1, 52-62.
Burgess, A. E. (1990) High level Visual Detection Efficiencies, in Vision: Coding and Efficiency, Ed. Blakemore, C., Cambridge University Press, 1990.
Cairns-Smith, A. G. (1996) Evolving the mind: on the nature of matter and the origin of consciousness, Cambridge University Press
Chalmers, D. J. (1995) Facing up to the problem of consciousness, Journal of Consciousness Studies 2:200-219
Chalmers, D. J. (1996) The conscious mind: in search of a fundamental theory, Oxford University Press
Crick, F. (1984) Function of the thalamic reticular complex: the searchlight hypothesis. Proceedings of the National Academy of Sciences, USA, 81, 4586-4590.
Crick F and Koch C (1990) Towards a neurobiological theory of consciousness. Seminars in Neuroscience, 2:263-275.
Crick F.H.C. (1994) The Astonishing Hypothesis: the scientific search for the soul , Simon and Schuster.
Damasio, A. R. (1989) The brain binds entities and events by multi-region activation from convergence zones, Neural Computation, 1, 132-132.
Damasio, A. R. (1994) Descartes' Error: emotion, reason and the human brain, G. P. Putnam, New York
Dennett, D. C. (1991) Consciousness Explained. Boston: Little, Brown
Fröhlich, H. (1968) Long range coherence and energy storage in biological systems, Int. J. Quantum Chem. 2, 641-649
Grace A. (1996) Thalamocortical Aspects of Consciousness from the Perspective of a Neurobiologist, contribution to the ASSC electronic seminar on Thalamocortical foundations of conscious experience, http://www.phil.vt.edu/ASSC/ esem.html
Gray, J.A.. (1995) The contents of consciousness: A neuropsycho- logical conjecture. Behavioral and Brain Sciences, 18:4, 659-722.
Grundler, W. et al (1982) Nonthermal resonant effects of 42 GHz microwaves on the growth of yeast cultures, in Coherent Excitations in Biological Systems, H. Fröhlich and F. Kremer (eds) Springer-Verlag.
Hameroff, S. R. (1994) Quantum coherence in microtubules: a neural basis for emergent consciousness ? J. Consciousness Studies 1, 91
Hardcastle, V. (1994) The binding problem and possible solutions. Journal of consciousness studies 1, 66
Hardcastle, V. (1996) The binding problem. Symposium on "Synchronous Oscillations and The Emperor's New Clothes" (available at http://www.zynet.co.uk/imprint/ online/hard.html).
Hayes, P. (1985) The second naive physics manifesto, in Formal theories of the commonsense world, J. Hobbs and R. C. Moore (eds), Ablex
Jones, E. G. (1985) The thalamus, Plenum, New York
Koch, C. and Crick, F. (1994) Some further ideas regarding the neuronal basis of awareness, in Large-Scale Neuronal Theories of the Brain (Eds C. Koch and J. L. Davis), MIT Press, Cambridge, MA
Llinas, R. R. , Ribary, Joliot and Wang (1994) Content and context in temporal thalamocortical binding. In G. Buzsaki et al. (Eds.) Temporal Coding in the Brain. Berlin: Springer Verlag.
Lockwood, M. (1989) Mind, Brain and the Quantum Oxford:Blackwell
Marr, D. (1982) Vision, W.H.Freeman
Marshall, I. N. (1989) Consciousness and Bose-Einstein condensates, New ideas in psychology, 7,73-83
McDermott, D. (1986) A critique of pure reason, Yale University report
Mesulam, M. (1985) Principles of behavioral neurology, Philadelphia, Davis.
Newman, J. (1995) Thalamic contributions to attention and consciousness, Consciousness and cognition 4, 172 - 193
Newman, J. (1996), Lead review and summary for the ASSC electronic seminar on Thalamocortical foundations of conscious experience, http://www.phil.vt.edu/assc/newman/
Newman, J. and B. J. Baars (1993) A neural attentional model for access to consciousness: a global workspace perspective, Concepts in Neuroscience 4, 255 - 290
Penrose, R. (1994) Shadows of the mind, Oxford University Press
Scheibel, A.B. (1980). Anatomical and physiological substrates of arousal: A view from the bridge. In J.A. Hobson & M.A.B. Brazier (Eds.), The Reticular Formation Revisited, pp. 55-66. New York: Raven Press.
Sloman, A. (1975) Afterthoughts on analogical representations, in Proc. theoretical issues in natural language processing, Cambridge, Mass.
Stephens, D.W. and J.R.Krebs (1986) Foraging theory, Princeton University Press
Steriade, M. , E. G. Jones and R. R. Lllinas (1990) Thalamic oscillations and signalling, Wiley
Taylor, J. G. (1991) Can neural networks ever be made to think? , Neural Network World, 2, 797-812
Taylor, J.G. and Alavi, F.N. (1993). Mathematical analysis of a competitive network for attention. In J.G. Taylor (Ed.) Mathematical Approaches to Neural Networks. Elsevier Science Publishers B.V., 341-382.
Tilley, D. R. annd Tilley, J. (1990) Superconductivity and Superfluidity, 3rd edition, IOP
Weld & de Kleer, J. (1989) (Eds.), Readings in Qualitative Reasoning About Physical Systems, Morgan Kaufmann.
Wigner, E. P. (1961) Remarks on the mind-body problem, in The Scientist speculates, Princeton University Press
Worden, R. P. (1995) An Optimal Yardstick for Cognition, Psycoloquy, 7 (electronic journal)
[1] The master map diagram of figure 2(a) is simplified for clarity. It is not meant to imply that there are no connections at all between the `satellite' sensory maps; just that the dominant connections, for the purpose of building a self-consistent model of local reality, are to a master map.
[2] Because the visual field only has high resolution near its centre, we might expect the required resolution of the 3-D spatial representation to be less than one part in 103 for some regions, leading to a total storage requirement of less than 109 ; nevertheless the requirement is large, and I shall continue to use the figure 109 for illustration.
[3] One alternative is that the excitations span the whole volume of the brain. Because of the large-scale inhomogeneities in the brain, this seems less likely.
[4] Each piece of visual input must be placed quite precisely in represented space, requiring a precise corticothalamic steering signal; thus corticothalamic fibres outnumber thalamocortical fibres by about 10:1 (Newman 1995)
[5] While some of this can be remembered (or reconstructed) sense data, the most vivid consciousness comes from current sense inputs. Even abstract and verbal thought is made conscious through the sounds of words heard in the mind's ear, or as mental images of metaphors.
[6] Since the thalamus is a bilateral organ, so why are there not two consciousnesses? Possibly there are, but in normal subjects we would expect the two to be very well coordinated with one another - so well coordinated that the differences are negligible. The observations of split-brain patients are relevant to this issue.
[7] Having several different types of excitation is is not contrary to the identical-particle nature of a Bose condensate (e.g both phonons and quantised vortices can exist as distinct excitations of superfluid Helium, supported by the `substrate' of identical Helium atoms).