In the endless search for self-knowledge through the centuries a central question has been how we come to have any knowledge at all. Common sense has no doubts: as John Locke put it “perception is the first step towards knowledge, and the inlet of all the materials of it.” But if this is so where does the news come from, and how and where is it received? Does it arrive in ready-made packets at some place in the brain where it is collected by a central agent, the mind, soul, or some hidden homunculus? Worse still, what is it that perception receives? Where do the facts come from? Are they all out there in the form of objects and events in a real world?
These are classical questions for philosophy, but recent approaches to them have included studies of the brain by neuroscientists and the construction by cognitive scientists of theoretical and practical models of perception and thinking. The theory of neural Darwinism as proposed by Gerald Edelman of Rockefeller University brings these two approaches together. He gives a detailed account of brain activity, and provides computer models of the principles that are involved. He claims to show in this theory how selection under the influence of experience is the principle by which brains accomplish their feats of perception, learning, and memory. Edelman wisely leaves to later the application of these ideas to the general problems of consciousness, emotion, and language.
In spite of this restraint the theory does in fact give promise of providing understanding of some problems that interest us very much, such as how there can be recovery of function after damage to the brain. The theory also provides a view of how we come to analyze the world, without postulating the presence of a non-material central agent or homunculus. This is a problem that is rarely faced by neuroscientists, who may study nerve cells, conditioned reflexes, cortical areas, or brain waves without deciding who or what it is that “uses” these to produce behavior, much less mentality. In spite of a vast amount of new information there is still no generally accepted view of how the brain functions as a whole, and most physiologists would probably say that it is not realistic even to think of such a thing. The very idea smacks of the “homunculus fallacy.” Apart from metaphysical considerations, anatomists and physiologists do not find any evidence for central agents, so they are left without any general theory at all.
Perhaps the reason for this serious failure is that neuroscientists mostly think about individual units but not about the population of neurons and their connections, the synapses in the brain. Populations are the very subject matter of Edelman’s theory of neural Darwinism. As the name suggests this theory holds that the organization of the brain is achieved by competition for survival between neurons, and hence by selection of the nervous connections that are effective for survival. Nature does not produce its marvelous results by means of precise machines such as an engineer would try to design. Rather, as Darwin showed, nature starts out with large numbers of varied individuals so that some will be suited for whatever the environment throws up. This method is not inefficient, as it might seem to be to an engineer; on the contrary, though expensive, it is the only possible way of providing for many unforeseen eventualities.
In the book under review Edelman shows that the principle by which nervous systems are built is to provide numerous sets of neurons with varying connections, from which those that are used are selected to make a system that ensures adequate responses to the environment. The processes that take place in the embryo do not produce a precisely connected brain in which every detail is specified by heredity; to make the necessary multimillion specific connections would require a magical micro-electrician, “the second cousin of the homunculus” as Edelman calls him. In any case no one has been able to show that any such precise connections exist, nor that they could provide the extraordinary perceptual capacity of a brain.
What has to be explained is the fact of generalization, that is to say, the power to recognize classes of things under many different conditions. Animals and people can recognize objects in all sorts of aspects after they confront only a few examples of them. Edelman cites the case of the pigeons trained by Richard Herrnstein and J. Cerella to peck at a key when photographs of particular objects were shown. They could learn to peck at any picture containing a tree or even a leaf and would not peck at a picture with no trees. This was perhaps to be expected, since trees are important for birds. The surprise was that the pigeons could also be trained to recognize pictures of fishes under water, in all sorts of variations of color or distance. Even more dramatic, they learned to peck at pictures of a particular woman in all sorts of clothes and rejected another woman in the same clothes and on the same street.
The problem is to discover how a brain is produced that is able to perform such feats of classifying and recognizing features of the world that are significant for it although they do not appear already labeled. It is clear that the genes cannot organize every connection; there are not nearly enough of them to specify the huge number of 1014 synaptic connections in the human brain. So some self-organizing process must be at work as the nervous system develops. The young growing nerve cells make fine filaments, pushing out like roots among the other tissues of the embryo. The directions they take and the connections they form depend on the properties of their surfaces. Edelman and his colleagues have discovered certain chemicals that make the growing nerve fibers attach to other nerve cells and are therefore called “cell adhesion molecules” or CAMs. The CAMs are produced by the surfaces of the nerve cells and fibers, under the influence of their genes. Several sorts have already been identified chemically. Their effect is to guide the growing nerve fibers so that they produce small interconnected groups of nerve cells in the embryo brain. Then, after birth, as the effects of stimuli begin to pour into the brain from eyes, ears, and skin, there is competition for survival among these groups of nerve cells. The cells and connections that are used survive, and the unused ones die away.
If the brain is to guide the creature accurately through the world there must be some means of ensuring that its calculations refer to the actual configuration of events. There must be a system of signposts providing a map that maintains all actions in true orientation with the world. This necessity is met at an early stage of development by the formation of maps in the brain. The best known of these maps lie in the visual system, where all the points on the retina are connected to neighboring points in the brain. This provides the animal or person with information about the spatial distribution of events, to which all visual perception can be related. Of course there are also gravity receptors to ensure that the eyes are held in their correct positions. Every animal that has an elaborate brain possesses such maps; they are found in man and monkey, frog, fish, octopus, and bee.
How can such correct connections be made? Various microhomunculi have been suggested; in each case it is postulated that tagged nerve fibers from the retina reach to correspondingly marked targets in the brain. But experiments such as those of Michael Gaze in London, mostly with frogs, show that there are no such tagged fibers. The correct connections are made not in that way but by a dynamic process of competition. What happens seems to be that the earliest fibers from the retina arrive guided by their CAMs, and their branches then spread widely in the brain so that their endings at first overlap. Some endings of each fiber make connection with the young brain cells, forming a coarse map of the retina. Then, as vision begins and the retina itself grows larger, further fibers arrive and there is competition for connection with the cells. Little is known about the details of this struggle but it probably depends upon visual function: those connections that are used become proportionately strengthened and the less-used ones wither away until a precise matching is achieved.
This system of competition and selection of connections, Edelman proposes, is the principle of all later development in the brain. It is an active–dynamic, adaptive system, finding the right result by trial and error rather than by a fixed set of rules. The system depends upon the initial formation of a large number of cells and branches, and then development only of those that are used. It has long been known that many nerve cells die during development, as many as 70 percent, and this expenditure is of course just what would be expected in a system based on competition and selection.
The maps of the sensory fields are the first stage of analysis. They become connected with further areas of the brain in which the perceptual process is carried further. S. Zeki of the University of London has found many such areas for vision in men and monkeys, and each of these parts of the brain serves to extract some feature of the scene, say contours, or colors, or distances. The formation of such a system must depend first on the genes, operating by CAMs, which guide the young fibers. The later stages are then produced by competition among the groups of fibers that pass from each area to the next.
So the early processes of development produce large sets of neurons of slightly varying forms in different parts of the brain. As the young creature goes about its business the sense organs provide signals that make connection with certain groups of cells and strengthen the connections between those that are activated together. Edelman shows that, given certain plausible assumptions about how the cells either stimulate or inhibit growth, the groups would become organized to make maps corresponding to the sensory surfaces, say of the retina or the ear or the surface of the hand, such as are known to be characteristic of the cerebral cortex. The brain thus comes to contain a model of the world, as it were, an idea suggested as long ago as 1943 by Kenneth Craik of Cambridge. As developed by Edelman the idea becomes a plausible reality.
The system of selection from a set, followed by amplification, is also the mechanism by which we acquire immunity. Antibodies are produced by cells selected from a large number of slightly different possible producers of them. It is probably no accident that Edelman, before taking to neuroscience, won a Nobel Prize for his studies in immunology; but he wisely insists that neuroscientists should not indulge in glib analogies with the detailed mechanisms of natural selection or clonal selection in immunology. It is the principle that is important. The great power and flexibility of a system built in this way depend on the fact that the groups formed under selection have overlapping functions. Each group responds mainly to one feature, say a visual contour set at a particular angle, but it also responds to others, and vice versa; each feature stimulates many groups. This characteristic, known as degeneracy, also explains an organism’s great adaptability and capacity for recovery after injury. There is increasing evidence, such as that of P. Wall’s experiments on cats and more recently M. Merzenich’s on monkeys, that after injury, say by loss of a limb, there is quite a rapid reorganization of connections in the brain. This potentiality for rearrangement may be important in many circumstances and provides hope for recovery after a stroke or injury to the brain.
The speed of changes that can occur in the brain after injury shows that there exist many possible connections waiting to be used if needed. If this is so in all parts of the brain then this gives us the capacity to train ourselves in new skills of action and indeed of thought. There is evidence that the power to make some types of new connection is limited to critical periods in early life—for instance when learning to speak. Yet there is no doubt of the existence, even in adults, of the potential connections that were shown by the work of Wall and Merzenich. The general conclusion is that the functioning of the brain depends upon the way that it is used and the fact that capacities not developed will atrophy. Human experience has long suggested something of the sort and it is good to have material proof of it.
The primary attraction of the selectional theory is that it makes sense of known features of brain anatomy, including the large numbers of cells and their connections and their variation, and the early death of many of them. No single experiment can be expected to prove or disprove such a theory. Edelman, however, provides computer models that show that it can work. His model of the part of the cortex that receives stimulus from the skin of a monkey’s hand develops a pattern of cells very like that which is actually present. Moreover the model adapts itself after the loss of a finger or toe just as happens in the cortex of a real monkey. A model of an automaton that Edelman calls Darwin-Wallace shows a projection from a sense organ to a series of reciprocally connected maps such as occur in the visual cortex. Darwin-Wallace learns to recognize letters even when they are of different styles or greatly distorted; that is to say the model performs the feat of generalization.
These models do not of course prove that the nervous system works in this way; but the groups of neurons and the maps that the models simulate are present in the brain. The assumptions that are made about the variability of neurons and methods by which they become connected are realistic; that the model can take account of the variations is the secret of its success. But the variations would pose an impossible handicap for a machine using a precise algorithm.
The selectionist hypothesis thus shows how it is possible for effective perception and generalization to occur without a homunculus. But the success of the model has been checked by human recognition. To get rid of the homunculus altogether would require a model that learns to act in a manner appropriate to what it perceives. Edelman does this by bringing in systems for the control of motor action according to the wants or needs of the individual. It has been known since the work of Sherrington that there are maps of motor control in the brain, with larger areas for parts that make more complex movements, such as the hand or tongue. Moreover the maps are very variable; an electrical stimulus applied to one point on the brain does not always produce exactly the same movement. Once again we see that the system involves maps with groups that are overlapping, degenerate, and variable, just as in the sensory areas. This is no doubt the basis for the great adaptability of actions. After any injury to a limb the muscles are at once used differently—as everyone knows when they limp.
These familiar facts agree with Edelman’s hypothesis of a brain formed by adaptive selection from sets developed by heredity. Then, throughout development and later learning the perceptual and motor maps become linked with the “hedonic” centers of pleasure and pain. It is significant that some of the most striking defects of learning and memory in humans are not from damage to the perceptual maps but from injuries to the temporal area at the side, where the sensory systems connect with those parts of the brain that are known to be active in ensuring the provision of the means of survival: food, warmth, sex, and freedom from fear. These “hedonic” areas somehow provide the signals that insure that the connections made in the sensory and motor areas are useful or “adaptive.” Unfortunately no scientific accounts agree on exactly how these connections come to be made.
In this sense the homunculus evades us still. He is hiding there somewhere between the sensory, motor, and hedonic centers; value is his last card. As the Cambridge philosopher G.E. Moore put it in 1903, “Good is good and that is the end of it.” The neuroscientist thinks that perhaps there is rather more to it than that. These connections should soon become known; we at least begin to see what to look for. For the present we may suppose that during development, with the same competitive mechanism, there develop connections between groups of neurons in the perceptual and motor centers, monitored throughout by signals from the areas “in the function of which [as Edelman puts it] certain evolutionary determined values, usually related to consummatory activity or fear responses, are embedded.” The problem of understanding learning is to find out how this monitoring by value is done, and of this we are still largely ignorant.
Perhaps even this scanty sketch may give some idea of the power of the theory of neural selection. Some physiologists may object that the theory is vague and not proven, others that the data are wellknown and the conclusions obvious; but no one before Edelman has put these facts together so convincingly. At a recent International Brain Research Congress in Budapest, Edelman showed a selectionist model more powerful than that presented in the book. It was able to recognize the shape of objects and to handle them appropriately. Such a system can be said to learn to perceive and act in an adaptive way; it does not need a homunculus.
It is not yet clear how far such systems based on random elements can replace current robots that use precise algorithms; they are certainly much more like the actual human brain. It may be that some of the very best systems do not start with complete randomness. Edelman himself quotes the evidence of P.D. Eimas and others that human infants are born with capacities for recognition of speech sounds. The characteristics of the earliest groups of nerve cells must indeed determine the performance of any brain and they are set by heredity. Some of these elements may already be tuned to detection of complex features. For example, there is good evidence from the work of Rolls and Perrett in Oxford, that adult monkeys have nerve cells that respond when the animal sees a particular face. Perhaps the primary set of inherited cells in a monkey (or human) includes some that are especially tuned to learn to respond to such features as language and faces that are so important to the social life of a monkey, or a human being.
In an early version of the theory of neural selection and amplification I suggested that simple learning in an octopus depends upon selecting between possible outputs from performed cells able to detect the direction of contours. Numerous highly oriented cells are indeed present in the optic lobes of octopuses. It seems likely that many animals are born with some such systems that are preset to learn in ways that give powers of perception appropriate for life, just as our system seems to be prearranged for speech. But in all animals that can learn anything beyond the simplest conditioned reflexes the brain has huge numbers of cells and connections. Even if some of them are preset to help learning in certain useful ways it seems certain that these populations will show the variety and hence the adaptability that is so valuable and that the theory of neural selection stresses.
Edelman’s book covers far more in theory and practice than it is possible to review here. Without going into mathematical detail he outlines a theory of great richness and power. It is characteristically what the philosopher Daniel Dennett has called “centralist,” in contrast to the “peripheralist” theories of stimulus-response behaviorism. It treats the organism as a developing agent with certain internal powers and tendencies to self-survival, but without any central controller. Every creature is unique and acquires its characteristics by interaction with the environment that it meets.
Edelman makes no grandiose claims, but it is interesting to consider how the theory affects our thinking about the possibility of conceiving a man-like artifact. Why do we find it so difficult to imagine a machine that thinks or believes or hopes? After all, the human beings who do these things are material objects; what is so special about them, other than language and consciousness? Two things stand out: first, each human being is the product of the long, continuous history of change during evolution and in his own life. Second, each human being shows an immense detailed knowledge of the world in all its aspects. We cannot yet make a copy of evolutionary history, but with a developmental theory such as Edelman’s we can just imagine a selectionist machine that, during its prolonged life, might gradually acquire enough experience to generalize about the properties of the world and, as a result, to show evidence of hopes and beliefs about the future. Who can say whether it would be conscious or show the properties of “intentionality” that are considered by some philosophers to be the hallmarks of humanity?