Content and Consciousness

by Daniel C. Dennett

  We need not reach any decisions regarding the disagreements in the field over the mechanics of cerebral plasticity or transmission. The role of RNA changes in the cells, the chemistry of the synaptic crossing, the generalized effects of drugs on plasticity, need not concern us provided that our general features and principles are embodied one way or another in the brain. There are other elaborations and complications of this picture that are more relevant to our general outline, but in the interests of maintaining our conceptual account as much as possible invulnerable to empirical disconfirmation in the laboratory, these will be ignored. For example, it is tempting to suppose, and there is some evidence for supposing, that particular synapses that regularly contribute to firings of a neuron tend to lower their frequency requirements, perhaps accomplished by a narrowing of the synaptic gap due to stimulated growth of the endbulb and dendritic knob, but it is not essential that we suppose this. There is one other feature of the brain’s physiology which it is important to mention, not because it is a feature required by our particular hypotheses, but because it is a feature required of all reliable information processing systems, and that is the brain’s use of redundancy and the ‘ambiguity’ of neuronal ‘signals’.

  No sense has yet been given to the claim that a neuron’s impulses are signals with content or meaning, but if, for example, a particular neuron in the optic nerve fires its output if and only if there is a particular pattern of stimulation on the retina (due to the particular summing effects of the neurons in the lower ranks leading to its input), in a borrowed sense one could say that the neuron’s output is unambiguous. However, except at the most peripheral levels, neuronal firings turn out to be, in this sense, ambiguous. That is, a wide variety of different, in fact very dissimilar, stimulus patterns may cause a neuron to fire, so that its signal is highly ambiguous. This fact, distressing to the neurophysiologist intent on ‘breaking the neural code’, is vital, however, to the successful functioning of the brain. The brain, for all its occasional lapses, is a highly reliable organ; seldom if ever does a complete failure of stimulus interpretation occur. If each neuron had only a function, and this function was not duplicated by other neurons, the death or malfunction of any neuron would throw all that followed into disorder. At the peripheral level – near the retina, for example – the death of a neuron might only cause a small ‘blind spot’ or imperceptible loss in colour discrimination or something of the sort, but if a single neuron at a high level were to carry single-handedly some information about a highly complex pattern of stimulation, its breakdown would cause something like total blindness for particular shapes or wildly mistaken identification of objects in the visual field. Neurons do not regenerate like other cells, and their mortality rate may be in the neighbourhood of one neuron a minute. Neurologists estimate that random malfunction of about one per cent of the neurons in the operation of any brain tissue or structure is normal. Clearly the reliability of the brain is greater than that of its components. Arbib presents a calculation to show the effect of random failures:

  Consider a chain of n modules [neurons] and assume that there is a probability p of malfunction for each neuron. Then the probability that the output of the chain is correct is, to a first estimate, (I - p)n. Now no matter how small p is, (I - p)n gets to a ½ when n is made large enough and if our output is equally likely to be right or wrong, it is of no use to us!5

  He goes on to point out that if p is one per cent, a neuronal chain of only 70 elements will have a probability of correctness of ½, and 70 elements is not very deep for the human brain, with its 1010 neurons.

  Reliability of transmission using unreliable elements can be achieved provided there is signal duplication in some form. If, for example, a ‘message’ is transmitted simultaneously by five neurons, and the probability of successful transmission for each neuron is high, say 0.99, the probability that successful transmission will occur in at least three channels is much higher. Then, if a statistical or vote-taking mechanism is inserted between each level of transmission, random errors due to malfunction will be absorbed as soon as they occur. The variable threshold capacity in the neuron can perform this function, provided the redundancy of signals is great enough, and provided there is a rich enough interconnection of outputs with next-level neurons.

  Simple redundancy, however, with each neuron’s output serving one purpose, would require an inefficient multiplication of elements. If, on the other hand, the signals fired by each neuron are ambiguous (as they are), if each neuron contributes to many different multiple transmissions, redundancy can be achieved with less elements. It is then the more or less simultaneous concatenations of neuronal outputs or signals that are unambiguous, rather than the outputs of individual neurons. The convergence of different concatenations of ambiguous signals at each succeeding level would partly resolve the ambiguity just as the convergence of ambiguous definitions determines unique or nearly unique solutions to crossword puzzles.

  The crucial point that emerges from this is that the candidates for vehicles of content or significance in the brain are compound. Afferent-efferent functional structures, which are to be sorted according to their appropriateness, have parts and could be ‘rebuilt’ piecemeal under certain conditions. The features of variable threshold and compound ‘signals’, together with the hypothesized initial situation in the brain of rich afferent-efferent interconnection and some partial appropriate pre-wiring, provide the elements needed for a hypothesis of evolution in the brain capable of explaining the brain’s ability to discriminate by significance and store and use information intelligently.

  The problem set up earlier was how the brain could cull out the appropriate afferent-efferent connections from the initial abundance of haphazard connections, but given the compound nature of neural signals we can no longer look for there to be whole compounds among the initially senseless fabric that are appropriate. That is, the odds are certainly against finding fortuitous structures sufficiently large and complex to produce or direct anything as sustained as a bodily motion would have to be, to be a demonstrably appropriate response to the stimulus environment. Where no connections would qualify as appropriate, how is selection to proceed?

  The answer comes from taking another hard look at evolution of species. For there to be evolution, there must be conflict between some features in the environment and the species to be eliminated. The only way any functional structures could be sorted within the brain would be if some of them were to conflict with the pre-established, wired-in, appropriate connections. There must be conflict and something must give. Clearly what must stand firm are the inherited connections. No other conflict, and no other outcome of the conflict, would resolve itself along appropriate lines. The inherited wiring or programming must be granted hegemony in all conflicts if the plasticity of the brain is not to undo the work of species evolution and leave the animal with no appropriate responses at all.

  So long as the other initially salient neural pathways are haphazard and uncompounded into ‘coordinated’ functional structures, what sense could be given to a notion of appropriateness or inappropriateness of these connections? All structures, it would seem, are going to be equally neutral in this regard. One bit of babbling or finger-twitching is no more or less appropriate than another. Some such structures, however, might conflict internally with the pre-wired connections, and although these would not be environmentally inappropriate structures by themselves, they would stand in the way of the completion of the pre-wired connections. These inherited links must, in addition to stimulating certain muscles in a certain sequence when presented with certain stimuli, also block the stimulation of conflicting muscle motion and the perpetuation of any neural structures that would in any other way interfere with the operation of the inherited links. Since for any afferent-efferent functional structure to become genetically established it must be environmentally appropriate over the long run, and since for any such structure to be appropriate it must be capable of survivi
ng in a plastic brain, all genetically established afferent-efferent structures must have, in addition to the appropriateness of their unimpeded function, the general capacity to inhibit competing connections. If this is the case, any of the initially haphazard connections that inadvertently competed with pre-wired connections would be inhibited and eventually, through inactivity, become inoperative, while any compatible haphazard connections would be allowed to complete themselves, and, by our principle of propagation, they would tend to recur. Just as in species evolution, it is thus not death itself that extinguishes a species, for all animals die and all neural events come to one end or another, but the failure to reproduce. This would have the effect of pruning the initially unstructured connections along lines at least compatible with and occasionally contributory to the appropriate inherited links already endorsed by species evolution. It would allow for the reproduction of everything that is at least not inappropriate, taking whatever inherited links there are as the arbitrary but contingently accurate standard of appropriateness.

  Harlow uses the term ‘baroque’ to describe those features that become genetically established through natural selection and exceed the functional, and this capacity for the propagation of the baroque is essential to the evolution of many capacities found in nature. Wings, for example, could not evolve fully developed in one fell swoop, and yet until they are fully developed, they have no positive survival value. The ability of a species to maintain through generations a fractionally or potentially appropriate feature is the sine qua non of complex capabilities and structures, and this holds particularly true for the obviously enormously sophisticated structures that must be required to control the behaviour we observe in animals and human beings.

  The gradual effect of this gentle sorting action will be to establish new functional structures, but if these are to have any permanence they must similarly be capable of overruling competitors, although here which competitors overrule which will not be entirely a matter of precedence of establishment, as it is with the pre-wired structures, for we want to allow for the unlearning of behaviour that eventually turns out to be inappropriate. As relatively permanent new structures are laid down, the efficiency of the sorting action will, of course, increase. An early manifestation of this evolutionary pruning will be the gradual smoothing out of bodily motions into more coordinated and graceful motions, and the resulting locomotion, the capacity for which is built up piecemeal, will bring to the animal new ‘experience’ in the form of novel stimulus patterns. These in turn will ensure that a constantly changing and novel afferent input will be presented to the brain (the analogue of mutation in species evolution) and the efferent continuations that happen to result from these new afferents will in turn be sifted. The effects of increasingly appropriate motion include an improvement in the quality of information brought in by the afferent barrage, as appropriate efferent structures controlling the focusing of attention, opening the eyes, and so forth become established. (There is a good deal of evidence that the controls for maintaining steady eye position, focusing, and maintaining standard orientation of retinal images are genetically transmitted, thus ensuring from the very beginning some regularity in the afferent barrage from the eyes, but the extent of this is not important here.) Thus the process is a repeated self-purification of function, gaining in effectiveness as more and more not inappropriate structure becomes established.

  Intuitively, the speed at which the evolution takes place will depend in part on the extent and rigidity of the initial programming or pre-wiring, and this is borne out in nature. Many animals are born with mature capacities for locomotion and discrimination of objects in their environment, but the greater the initial ability, the more rigid the brain, and hence the less adaptable the animal. More intelligent animals require longer periods of infancy, but gain in ability to cope with novel stimuli because of the higher proportion of ‘soft’ programming – programming not initially wired in and hence more easily over-ruled by novel stimuli. The speed of evolution is in any case incomparable to the speed of species evolution, for the counterparts of generations endure not for decades or months or even minutes, as in the case of some primitive organisms, but for a few milliseconds.

  If it be doubted that such a slight force could account for the learning capacities of animals and men, the fact that species evolution has produced ‘instinctual’ behaviour in some animals the equal of learned behaviour in others is some support for the claim. Babies must learn to see and walk, but whatever controls this in babies has a counterpart in chicks and puppies, and in these creatures the controls are clearly almost entirely inherited. Species extinction is as slight a force as the extinction through incompatibility posited for the learning brain, and yet species extinction has been a strong enough force over the years to produce such complex behavioural controls as those governing the ‘territorial’ behaviour of some birds, food discrimination, and specific patterns of defensive behaviour.

  Implicit in these arguments is a corollary to the effect that the Lamarckian hope that some acquired characteristics may be genetically transmitted is gratuitous. The fear that makes Lamarckian hypotheses attractive is the fear that species evolution by itself would not be effective enough to produce the sophisticated ‘instinctual’ behaviour observed in the animal kingdom, so individual acquisition of know-how is rung in to help. But I have argued that only an intra-cerebral evolutionary process could account for such individual acquisition, and if an intra-cerebral evolutionary process can produce sophisticated behaviour, it follows that over a longer run species evolution can do the same. The transmission of acquired characteristics is not ruled out by this argument; it is just denied the crucial role it might seem to have.

  The intra-cerebral evolution hypothesis also allows the controversies over instinctual behaviour and the interpretation of deprivation experiments to be seen in a new light.6 The standing difficulty with deprivation experiments has been the near-impossibility of so reducing the stimulus environment of the animal from birth that the possibility that the behaviour in question is learned can be ruled out. This has been a difficulty because some stimulation is always necessary just to ‘trigger’ the behaviour in question. The results of experiments have tended to blur the fine line between innate and learned behaviour that was seen as a desideratum. Does the animal have the particular behavioural capacity intact at birth, or does it have some inner state at birth which allows it to ‘learn’ the behaviour almost instantaneously when the right stimuli are present? The distinction loses much of its importance given the inchoate view of learning presented here. The existence of some degree of wired-in behavioural controls is established here not by the results of deprivation experiments but on conceptual grounds alone. That is, it is argued that without some such foundation for appropriate behavioural discrimination in the brain, the brain as a physical organ could not learn at all, since it would have, as it were, no ‘standpoint’ from which to make initial discriminations. The extent of pre-wiring in each species is subject to experimental determination, but there is no assurance that the exact limits of pre-wiring will be determinable via behavioural manifestations. That is, such basic pre-wired controls as those governing reflex withdrawal from painful stimuli have obvious behavioural manifestations, but much more sophisticated behavioural controls may be genetically transmitted and yet because they are only partial in their pre-wired form, deprivation experiments would not reveal their existence. It is likely, on this view, that partial appropriate afferent-efferent connections could be established the completions of which would have to be learned. There is no sharp distinction in efficacy between species evolution and individual intra-cerebral evolution, so where species evolution leaves off and intra-cerebral evolution takes over is a matter of no great importance as far as the survival value of the pre-wiring goes. A set of potential behavioural controls has virtually the same survival value as complete behavioural controls, given a regularity in the early learning environment of the specie
s ensuring the completion of the controls in most cases. Behavioural evidence for such partial structuring of the infant brain might be extremely indirect and not at all conclusive. For example, it might well be the case that in a human being there is a partial smile-discrimination mechanism, the completion of which requires a relatively large amount of learning, including at the very least the development of locomotion and sensory discrimination in general. The evidence for this is quite tenuous. Babies seem to respond appropriately to smiles very early, which is remarkable considering the complexity of what is communicated by a smile and the paucity of corroborating evidence in the environment for a smile’s significance. There can be nothing intrinsically friendly about the spatial configuration of a smiling face, and furthermore there must be a remarkable lack of uniformity in the retinal projections of different smiles. The universality of significance of smiles among the people of the world contributes to the suggestion that there is some partial discrimination mechanism genetically transmitted, and of course such a mechanism would have survival value since the early recognition of, say, parental approval or disapproval is a valuable capacity in the learning child – if not today, very likely in primeval days when an unlearned lesson could be fatal. (Survival value would depend, of course, on a concomitant inherited tendency to smile when one wished to show approval, pleasure, friendliness, etc. Alternatively, an inherited smile-recognition system could be entirely baroque or a no longer useful relic from man’s simian past.) Such a partial pre-wiring, if it exists, would not even come into play, would have no behavioural manifestations at all, until considerable learning had occurred, and hence would be a bit of inherited behaviour control quite inaccessible to testing by deprivation experiments.