The Location of Representations: Grandmother, Cardinal, Pontifical and Other Cells
Jonathan CW Edwards
Any theory that tries to integrate basic neurophysiology with real-life brain-dependent activities like parking a car is likely to invoke the idea of representation, in some form or other. To park, the car and parking space would seem to need to be 'represented' in some way in the brain processes that determine successful parking (Kosslyn, ). What physical processes might constitute 'representations' and what sort of location they might have are, however, less clear, at least partly because of ambiguities in the concept of representation itself.
The nature of representations
Representations might be defined simply as patterns of brain activity which arise in response to sensory signals encoding patterns in the world and which are specific to a particular external pattern. Thus, to a first approximation, a representation of a nine of diamonds card would consist of those activities in the brain arising in response to a nine of diamonds card that would differ if it were a three of clubs. The problem with such as definition is that the depolarisation of photoreceptors in the retinae is already such a representation and there may be at least a dozen further levels of representation in lateral geniculate bodies, right and left primary and secondary visual cortices, temporal, parietal and frontal lobes, en route to a behavioural response. All of these are likely to determine specifically the nine of diamonds and may do so in the absence of at least some of the others. Thus it may be appropriate to talk of many representations at many levels rather than a single representation.
Many representations at many levels may be a satisfactory account of the situation, and will be returned to later. However, as indicated below, there are reasons to be more stringent about criteria for what constitutes a representation. Moreover, at least within psychology, representations are often implicitly or explicitly considered to be associated with reportable experience. These may be a 'percept', as when something is viewed or heard, or a 'mental image', as when retrieving memories, thinking of a scene or sound, or, most vividly, in dreams and hypnogogic states. Moreover, there is a general assumption that there is only one instance of this type of representation at a time, and there has been debate in several different disciplines over the last forty years over whether such a single representation is local or distributed (ref). This debate remains unresolved, and it is suggested here that this may reflect a degree of confusion about what we should expect the biophysical processes underlying a representation, of the 'percept' type, to consist of and where they might be.
The purpose of this article is to explore the proposal that however difficult the question of precise location of the representations that underlie percepts may be it is a legitimate question and one that may be answerable, at least to a better approximation than we have at present. It may be necessary to use rather different methods from those of traditional neurophysiology to reach an answer, but having at least an approximate answer may be essential to the programme of understanding how neural connectivity generates intelligent behaviour. There is an alternative view that percepts are part of a 'functional' level of explanation that cannot be tied to a specific physical location. However, in all other contexts we consider that representations must have a physical location in order to be causally effective. (Since a 'functional description' explicitly invokes a causal chain, it is perhaps ironic that functional explanations often appear to be seen as not needing to be bound to any specific causal path.) Thus, to fulfil the original objective of linking neurophysiology to real life activity it would seem appropriate to maintain the idea that the representations that form the basis of reportable percepts have a physical location.
General neurobiological principles
Unless there are good reasons to do otherwise, an account of a physical location for a representation in a brain should follow general principles already established in neurophysiology. Two such principles are particularly relevant. The first is the neuron doctrine. The second is that the content of a percept will be encoded in signals that form inputs to the physical domain associated with the percept.
The neuron doctrine dates from the time of Ramon y Cajal and, in essence, is the principle that all brain function can ultimately be explained by the interactions of separate cellular units. For present purposes, it can be summarised as follows: Each neuron is a discrete computational (in the broad sense of something with rule-based input-output relations) unit, obeying, as a unit, classical biophysical laws. The timing of firing of a neuron is entirely determined by the interaction between chemical and electrical influences from immediately surrounding brain and the physical state of the cell. Firing can be affected by traditional synaptic transmission, by electrical effects at tight junctions, by diffusible factors with generalised effects on membrane potential and by aspects of the cell’s physical state such as temperature and metabolism, but it is not otherwise causally linked to activity in other cells.
It has recently been suggested that the neuron doctrine should be replaced by a description of brain function at a more 'global' level (Gold and Stoljar). However, since the causal biophysical pathways of the neuron doctrine are not seriously in doubt it is unclear that a more global description can be in any sense an alternative, rather than just an attempt to extract useful commonalities in rules for massively parallel events that may allow approximate accounts of overall dynamics. It is difficult to see how such commonalities are relevant to specific representational content, rather in the way that it is difficult to see how changes in times of postal delivery would influence the content of people's mail. In particular it is unclear that a desire to see ‘decisions’ as being ‘taken’ by groups of cells showing, for instance, synchronisation of behaviour, rather than by individual cells is based on any sound computational theory. A 'higher order' description of neuronal events in terms of systems theory or non-linear dynamics cannot contravene lower order biophysical laws. In simple terms, ten cells cannot be said to influence ten other cells in the way that ten men can lift ten rocks in a sling through pulling on a single rope. As far as we know there is no way to circumvent the causal bottleneck of the individual neuron (Barlow 1994).
The second premise is equally fundamental to neuroscientific explanation, but is rarely overtly articulated. It relates to Rosenberg's concept of receptivity (Rosenberg 200..) and is laid out in an explicit neurological context in a recent discussion of the basis of the qualia of experience by Orpwood. The representations we call percepts must be based on the availability of certain data to some neuron-based entity or domain, i.e. they must be the inputs to such a domain, which will also generate outputs in response that will allow the percept to be 'reported'. Something has to receive the signals that allow it to support a percept, whether these are derived originally from sense organs or other sources as in dreams.
This might seem self-evident. However, this second premise is worth emphasising because literature on perception and consciousness from disciplines outside neurology frequently appears to take a radically different view of the basis of percepts. Percepts may be seen as arising from, or being associated with, computational or ‘information processing’ operations, which involve not inputs but input-output relations - the essence of ‘functionalism’. This appears to imply that if percepts belong to physical entities then those entities are in some way privy to their outputs as well as their inputs. One of the most robust findings of neuroscience is that this is not so, and indeed in a sense it is self-contradictory. Moreover, as will be discussed below, if this involves access to the meaning of these outputs a paradox arises, since for a signal to have a meaning, as Orpwood points out, it must have a recipient to interpret it, for which it must be an input.
Representations are based on inputs
In other words, a representation must be a representation to something. Although it is commonly suggested that there are no 'inner receiving entities' for representations in a brain this is inconsistent with our understanding of causality. To be part of a causal chain, and thus reportable, the information encoded in a representation must be made available to something that generates a response. A word of text in a forgotton language embedded in an opaque medium that cannot be removed without destroying the text cannot usefully be called a representation. Similarly, a pattern of lines of cellular activity in the visual cortex of a monkey in its normal environment which bears a homotopic relation to a pattern of tree trunks the monkey is viewing is not in a meaningful sense a spatial representation, either to a scientific observer or to any part of the monkey since no part of the monkey, including the cells themselves, is informed of the spatial relations of cells with and without activity. The cells will represent sense data to other parts of the monkey's brain through patterns of downstream synaptic transmission, but the homotopy of the spatial relations of their cell bodies is of no causal consequence.
There is no doubt that treating representations as being based on inputs to some neural structure raises difficulties. However, available neurobiological evidence seems to indicate that our intuitive association of 'knowing' with an input to something is valid. It might be argued that this only applies at the peripheries of the system and that in between things must be viewed differently, but it is unclear why this should be so. Historically, neurobiological observation has suggested that we can push the concept of ‘input’ as far into the nervous system as interpretable observation will allow. Pressure from an intervertebral disc on a fifth lumbar nerve root gives pain in the big toe. Cochlear implants give severely deaf people an experience of sound. Stimulation of cerebral cortex in the awake individual can evoke sensations and memories. These observations indicate that sensory pathways, at all points up to that where a percept is experienced, are simply providing inputs to the next stage, that could be triggered artefactually without any change in the nature of the percept experienced. (Artefactual interventions are often associated with clues that allow us to know that the percept is aberrant, but it seems likely that in principle these could be avoided, as with an ideal virtual reality apparatus.)
The work of David Hubel (vision) and many others has shown that the process of acquisition and collation of sense data can be tracked quite far into the brain. Cells that respond to lines at particular angles, lines of limited length, or colour contrasts can be demonstrated. It might be argued that the absence of much analysis beyond this level could indicate that signals enter a ‘black box’ in which percepts are no longer associated with inputs, but rather with input-output relations. However, the simpler explanation is that beyond this level collation is so sophisticated that it is difficult to know how to set up an experiment to analyse it. The more recent work of Koch and colleagues showing that individual cortical cells respond to the faces of specific individuals suggests that this is so and that such difficulties can sometimes be overcome.
In summary, despite trends in other directions in some fields of study, the two assumptions of the neuron doctrine and the doctrine of percepts as based on inputs to perceiving entities appear to be on a sound footing.
Possible domains for representations
Armed with this basic logical standpoint, it is possible to ask some general questions about the location of the representations we call percepts and the nature of the entities to which these representations are available.
The starting premise is that our introspection tells us that at least one entity exists in a waking brain that has an experience correlated with the input from the sense organs of a certain human being. The validity of this could be denied, but to deny it would be a pity if this premise can yield certain badly needed clues to the way the brain dictates behaviour. To be useful we need to be able to describe such an entity in dynamic physical terms. The prima facie case is that it will consist of a dynamic domain in one or more neurons receiving inputs derived from all sensory modalities, and other internally generated signals, and sending a sequence of outputs to all, or most, motor pathways.
Following Orpwood’s exposition of premises relating to meaning, the next premise is that if the entity that has the percept is a domain in one or more cells, in accordance with the neuron doctrine the input will be of signals leading to patterns of electrical depolarisation of cell membrane which will, either directly or indirectly, be associated with an interpretation which is a percept. Since we are considering input this ought to be a pattern in dendrites afferent to the cell soma. Depolarisation of efferent axons will correspond to stereotyped output signals that cannot be expected to give rise to an interpretation in the cells in which they arise.
The term 'interpretation' may need clarifying here. It is not meant in the sense that sensory signals encoding four legs, a tail, pointed ears, red fur and a toothy snout are interpreted as a fox. That would imply at least one computation involving an input-output relation. Interpretation is used here to mean simply the correspondence of an input, (of electrical or chemical signals based on collation amongst sensory data and with data from memory) to a meaningful 'percept'. 'Manifestation' might be a better term, since it implies no computation, nor any physical interaction, but simply a correspondence between physical input and its meaning, or 'appearance', to the receiving entity. Interpretation of this sort must presumably occur at the final point of input since there is no means by which interpretation could be 'carried forward' from previous events in input pathways. (Since the 'cone' of past events contributing to any causal interaction is immeasurably complex the idea that 'interpretations' or 'manifestations' occurring earlier should be carried forward in causal chains is as absurd as it is inconsistent with experience.)
Representational domains cannot be based on traffic
The next step takes us beyond the initial assumptions in a rather more obvious way. It is that the percept must be an interpretation both of signals arising from membrane excitation and counterfactuals or ‘non-signals’ corresponding to sites where membrane might have been excited but is not at that time. Unless signals are interpreted in the context of all possible signals in a particular domain we lose any sense of information encoded in relations of signal to non-signal. A summation of all and only the black spots of a set of printed words has effectively only one meaning: black. Moreover, it becomes arbitrary whether the spots included are on one page, or in one book or in a whole library. Only if the white areas are included do we have diverse meaning and bounded domains of meaning. Thus, in vision, a 'line' of uniform colour within a block of uniform colour is not interpreted as a line. The interpretation of a line implies the absence of signals encoding similar colour on either side of the line.
This last premise has an important consequence. It means that the domain that supports a percept cannot be defined by a pattern of signal traffic; it cannot be defined in terms of where signals are occurring. It must include counterfactual non-signals, so there must be some intrinsically defined dynamic structural domain within which signals and non-signals are co-interpreted as percept. The domain receiving signals, interpreted as a percept, cannot be a ‘circuit’ or ‘network’ in the sense of a set of pathways currently carrying signal traffic. There may also be an important distinction here between the processing units in a brain and in a computer. In a computer there are ‘gates’ in which electrical signals ‘open’ or ‘close’ connections between units, forming and breaking electrical circuits. The brain does not have gates in this sense. Connections remain unchanged, at least over periods of hours, regardless of traffic. The processing units are integrators but not gates.
Problems with 'self-reference'
The implausibility of representational domains being defined by traffic patterns and the inability of domains to know their output, together raise difficulties for concepts of self-reference and its relation to perception. Orpwood makes the suggestion that a cell or group of cells that has an output that connects back to their input might support the qualia of experience because they are in some way self-referent. The problem here is that all cells within the cortex probably have the ability to indirectly influence their inputs through outputs and this must be considered just as much true if they influence their inputs by null outputs, as by substantive outputs. Thus the possibility of self-influence does not appear to help in defining perceptual domains. Moreover, since a domain cannot know its output and inputs do not provide an indication of their source of origin (as mentioned above) there is no way that a self-influencing domain can know that it is self-influenced, rather than influenced by, for instance, similar domains nearby, or at a mirror-image site in the other hemisphere.
Although the idea that re-entrant pathways might be crucial to perceptual experience is popular, these considerations cast doubt on such an idea in principle. It implies knowledge that is not provided by the available causal pathways and as such violates normal precepts of 'naturalistic' descriptions of the world. As implied by Hume, self-reference may be more of a virtual cultural construct than a true reflection of brain dynamics.
Localised versus distributed representations
The idea that the domain of a perceptual representation is defined by a structure rather than a pattern of signal traffic means that the issues about localisation and distribution of percepts raised by people like Lashley have to be taken seriously. Percepts, if in a degraded form, appear to survive damage to almost any part of, or any amount of, cortex. Removal of certain specific areas can produce predictable and well-defined defects, but does not appear to remove the capacity to have percepts, even if there is agnosia in the sense of not be aware that the percept is defective. The inference is that if the type of entity supporting the representations of percepts being sought is indeed cortical then there is no single local entity. That leaves the possibility of one very extended entity or multiple local entities.
The idea that a percept is an interpretation of the inputs to neuronal membranes over a wide area of brain generates a range of problems. Many neurones would appear to be involved in ‘housekeeping’ activity such as suppression of vision during saccades, early collation of sensory inputs, or motor activity. The inputs to such cells do not appear to figure in percepts; the biological advantage of percepts would appear to be that they reflect the input to a very select cell population involved in conscious attention. In this context, it is unclear why the inputs to certain cells and not others should figure in a percept. Nor is it clear why there should be a perception of a single copy of sense data when most if not all signals arising from cellular activity in sensory pathways are being sent to form inputs to many cells through widely ramifying axonal branches.
These and related concerns were probably the impetus to the proposal by Pribram (erf) that the cortex carries information somewhat in the manner of a hologram. In a hologram every part of a spatial array carries a version of the entire pattern of information being handled, but in a low-resolution format, with higher resolution emerging as larger proportions of the array are sampled. However, the problem with the holographic model is that it is not clear what the entity or entities with inputs corresponding to high-resolution percepts are to be. A simpler but related idea is that sense data are sent to many locations in the cortex and each of these has the potential to interpret its input as percept. This would seem to be in keeping with the experiments of Koch in which visual sense data often gave rise to excitation in many sampled cortical cells. In some cases cells were highly restricted in their responses to images, but others were more promiscuous. At least there is little doubt that sensory stimuli lead to signals being sent widely to many cells.
The above discussion suggests that the question of what domain supports a perceptual representation boils down to what it perhaps must always have been: a question of which cell or cells. Its formulation may, however, need to be slightly more subtle; How many of which sort of neuron have a perceptual representation encoded in their input(s) and do they constitute a single domain of a single representation of this type at any one time or are there multiple domains, supporting multiple representations based on the same sensory data? (The latter would not give rise to any sense of multiplicity since such multiplicity would not itself be represented or perceived by anything.)
Pontifical, grandmother and cardinal cells
The simplest hypothesis for the domain of a perceptual representation, now largely regarded as a null hypothesis, is that of the pontifical cell, as entertained in the nineteenth century and discussed by William James. The idea dates back at least to Leibniz, writing shortly after neurons were first observed. This form of pontifical cell is the single cell that supports all representational percepts of all sensory inputs. It is the 'me' cell. All other cells act as conduits of information to and from this central cell, collating inputs and delegating outputs. James considers that they might also support 'percepts', but of a meaner sort than that which we report as being 'mine'. The attraction of this idea is that a cell is a single functional unit, which has an intrinsically delimited input domain and human experience appears to be 'single'. However, the idea that one neuron should have this highly specialised function is implausible on a range of grounds and, as indicated above, the argument that experience seems to be 'single' is non sequitur, since there would be no reason for there to be representation (and thus perception) of multiplicity, or a sense of 'other copies', within each of multiple representations.
At this point it is useful to raise what may be a confusion in the perception literature between sites of representation and sites of recognition. (For the time being the discussion will be continued in relation to the domains of individual cells for ease of illustration of general principles.) Sherrington (1940ish) appears to have invoked a concept of a rather different sort of 'pontifical' cell in an attempt to explain recognition. Sensory data relating to an object such as a dog enters through many thousands of receptors. Recognition would appear to require sequential stages of discrimination, each leading to a reduction in the number of possible interpretations of the input pattern. This might be expected to take the form of a 'pyramid' with an order of magnitude fewer cells at each stage until the input finally converged on a single cell that had the specific responsibility of recognising a dog. In this case there would be a pontifical cell for a dog, another for a cat and yet another for a grandmother.
The important point here is that we have no particular reason to think that only the cell that has the job of recognising a dog will receive input signals relating to features of a dog. If a hundred cells each recognised a different mammal we would not expect the presence of a dog to lead to input to only one of these cells. We would expect all the cells to receive signals indicating features of a dog but only one (or some) to fire. It could be argued that synapses receiving signals encoding pointed ears will have been lost or rendered impotent on koala-recognising cells but the very function of a dog-recognising cell depends on a response to a specific combination of features none of which are in themselves specific for dogs.
Thus if a representation is based on an input pattern we do not expect sites of representation and recognition to be commensurate. This emphasises the need to consider a causal chain as potentially involving many levels of representation with multiplicity at each level. It highlights the fact that a representation is always a step in a causal chain and is thus always a representation to a domain at a particular point in that chain. Thus a pattern of data, perhaps encoding legs, fur and muzzle, would represent a dog to a 'dog-pontifical cell' as well as to a lot of cells tuned to other creatures. In turn the firing of the dog-pontifical cell and not of its neighbours would represent the presence of dog to the rest of the brain. The two types of representation would be entirely different. Moreover, intuition tells us that whatever domain has a percept of a dog of the sort normally discussed it must have an input encoding both the diagnostic features of a dog - legs, fur etc. - and the sense of these being part of a dog, apparently putting the relevant domain downstream of the site of dog-recognition with additional parallel input encoding the original upstream sense data.
It has become increasingly clear that recognition does not use a pyramidal system with fewer and fewer cells at each stage. Sequential stages involve as many, if not more, cells as at the beginning - rather as implied by the above discussion. Recognition is signalled by the firing of one or a few cells in the context of 'counterfactual' non-firing of many more cells. At all stages representations are thus widespread, but it needs to be established whether this is because individual representations are extended or because of multiplicity.
This issue is relevant to Barlow's classic 1972 paper in Perception (Barlow 1972). Barlow takes as his object a grandmother, following Jerry Letvin (196-) and discusses the likelihood of there being a 'grandmother cell' in the sense of a single cell that fires with 100% sensitivity and specificity for grandmother. This bears some relation to Sherrington's pontifical cell but not to that of Leibniz or James. Barlow suggested that grandmother was probably not important enough to have a cell to herself and that it was more likely that grandmother would be encoded by the activity of perhaps a thousand 'cardinal' cells, each representing an aspect of grandmother such as a mouth or nose, any of which might presumably contribute to encoding other faces in other combinations. These elements of the percept are then seen as combining rather in the way words combine in a sentence (an analogy also used by David Marr in 'Vision'). Note that Barlow is not proposing a redundancy-for-safety strategy with information being distributed in a 'holographic' way amongst several cells each with a sensitivity and specificity for a percept of less than 100%. He is giving each cell a separate and specific job.
The odd thing here is that Barlow appears to be describing the activity of cells upstream of a site of recognition of grandmother. If each of these cells is responding to signals which together encode a feature not entirely specific and sensitive for grandmother then grandmother can only be recognised, and appropriate social responses activated, by a group of cells receiving inputs from these thousand cells, some of which will fire and some not. It would be these next cells whose inputs would encode all grandmother's features and it would therefore be their domains that we could expect to support a 'percept' of granny in the sense of manifestation of all of grandmother's features, whether or not they fired. And it would be the pattern of firing and non-firing of these latter cells that would 'represent' to domains in the rest of the brain the presence, but not the pattern of features, of this individual. Whether or not within this latter group of cells there are cells with 100% sensitivity and specificity for grandmother is a different issue that does not necessarily bear on the search for the domains supporting the representations known as percepts.
The above discussion emphasises a number of issues relating to this key question. Representations in the brain must be multiple and diverse. At each level many cells will be involved in representation. Representation and recognition are not likely to be commensurate. So far the discussion has been in terms of individual cells despite the general assumption in the literature that representations in brains each involve at least a few cells and perhaps thousands. This issue needs to be revisited in the light of the preceding arguments.
A return to the neuron doctrine.
The term representation suffers from a number of weaknesses. It has already been noted that to be useful in the context of perception the term has to imply a step in a dynamic causal chain and its content ought to be encoded in the input of signals to some domain. It is also very difficult to see how a representation can have a meaning, or interpretation, of the sort associated with percepts, unless its content is encoded in the co-temporal input of a pattern of signals and non-signals to some domain.
It should be recognised that representations of this sort do not occur within computers. Stored data in a computer can represent something meaningful to an external operator capable of accessing it via a screen display but no representation based on a pattern of co-temporal input occurs to anything within the machine itself beyond the trivial input options to an electronic gate of on/on, on/off, off/on and off/off. Moreover, we have no reason to think that anything in a computer interprets, or attributes meaning, to any input signals, only outside observers need do that. It might be argued that a sequence of signals passing through a gate might constitute a representation to that gate. However, since each signal contributes to a separate computation this would appear to be invalid. There is no sense in which a sequence of incoming signals are subjected as a whole to a computation, other than as arbitrarily defined by an outside computer programmer. Within the machine any 'chunking' of serial signals into 'representations' adds nothing to the causal account and at the gate in question no such chunking would be apparent.
Within a brain there are units that do receive complex patterns co-temporally, in the form of individual neurons. Moreover, they are the only units that receive patterns as far as we now. Barlow's 1000 cardinal cells are not a unit receiving a pattern of 1000 features of grandmother. Each has a separate input encoding one feature. For all the 1000 features to contribute to a causally effective representation the 1000 cardinal cells must send signals that all converge on at least one neuronal unit downstream (well within the known range of neuronal inputs). If representations are to be based on co-temporal inputs and causally effective then they have to go through the causal bottleneck of the individual neuron, as indicted above. The neuron doctrine, as was probably evident to Leibniz, entails the very simple but surprising premise that representations in brains can only be in individual neurons. There may be very large numbers of such representations, all encoding the same sensory data, distributed over a wide area of cortex, but each representation must be tied to the receiving unit that is the neuron.
This conclusion immediately resolves the paradox of localisation and distribution of representation in the brain, since it implies that local representations are distributed widely. Thus although the phenomenon of representation is widespread in the brain any individual representation must be available locally to a single cell, just as the distribution of a newspaper is widespread but can only convey news if all the words of a news story are present in each copy read by an individual. To suggest that a representation can be available to several cells is equivalent to suggesting that the news can be understood by a group of people each of which only receives one word of the story.
The conclusion also resolves the question of precisely where in the brain are the representations that determine behaviour. The answer is that there are many all over the brain. Even the question of where in the brain are the representations that determine considered verbalised behaviour has the same answer. However, it seems reasonable to attach some special signficance to representations to cells with multimodal inputs that would allow both the visual and auditory features and the concept of a dog to contribute to a 'percept' of a dog.
Putting representations in individual cells may appear implausible. However, it is difficult to identify reasons why representations in individual cells should be more implausible with representations in many cells. The apparent problem may simply reflect the fact that if a specific location for such representations is proposed our lack of understanding of the rules of interpretation are brought into sharp focus. This may be no bad thing.
Another attraction of the idea that representations are to the single computational (rule based input-output) units that are neurons is that it implies that the brain does not perform operations on 'atomic' (structureless) symbols, but rather it performs operations on 'molecular' representations. That is to say that the basic data units that the brain operates on are irreducibly complex, with many degrees of freedom. This begins to address the bind of how the manipulation of symbols can be associated with an experience of complex patterns that reflect the complexity of their referents. It also provides a reason why, as appears to be increasingly recognised, syntax and semantics cannot be totally dissociated when considering meaning. The complexity of individual data elements allows computational rules to be content sensitive rather than purely syntactic.
Mainstream neuroscience prides itself in being rigorously physicalist, in the sense of adhering to the basic precepts of natural science and our understanding of causality. A consideration of representations in such a rigorous causal framework leads to the conclusion that all representations in the brain, including those that may form the basis of percepts must be based on the inputs to individual neurons. These representations will therefore occur at multiple levels of sensory processing and will be multiple at all levels, including the level attributed to percepts. Such a model is counterintuitive but resolves certain important problems relating to the distributed nature of representation and to linguistics.