Edelman’s Theory of “Neural Darwinism”
Edelman’s theory involves the following set of ideas and concepts:
1. During embryo development there is laid down an extremely complex neural network. This constitutes the primary anatomy of the brain.
2. Learning involves the superimposition of patterns of connections onto this primary network. Such patterns are not created by creating new, or more, connections, but by strengthening existing pathways.
3. Pathways compete with each other. Patterns of connections are nourished by stimuli and grow stronger at the expense of weaker ones, which decay and are overridden. Note it is the weaker connections which disappear, not the neurons themselves.
Let us now examine and expand on these propositions.
1. During embryogenesis there is laid down an incredibly complex neuronal network in the brain. Cellular adhesion molecules (CAMs) determine which cells attach to which other cells, thereby creating localized areas in the brain – the brain’s major substructures. These substructures include all the centres which analyze and code sensory inputs and motor outputs. Within these substructures of the brain, and across them, a myriad of connections are established. This is the primary architecture at birth.
Note that the combination of genetic and epigenetic factors which determine these connections (called synapses) are so complex that although the macro structure is roughly the same from one individual to the next (within any given species), the microstructure is infinitely varied and unpredictable. Genetic twins do not have an identical brain architecture at birth.
2. Learning involves the superimposition of patterns on this neural network. Such patterns are not created by creating new, or more, connections (synapses), but by strengthening existing pathways.
This idea – that changes I the strength of synaptic transmission due to usage of certain critical neural pathways underlies the learning process and represents the basis of memory – has been around for over 75 years. However, the first support of this hypothesis did not occur until 1973, when Bliss and Lomo found that the brief, repetitive activation of certain nerve pathways in the hippocampus of rabbits caused an increase in the strength of synaptic transmissions that could last for days or even weeks. Since the hippocampus is a part of the brain thought to be important in information storage, their findings appeared to be particularly significant. This long-lasting increase in the strength of synaptic transmission is called “long-term potentiation” (LTP). LTP represents the most likely explanation for the cellular mechanism underlying the phenomena of memory and learning […]
Thus the memory of an object such as a “locomotive” is not some visual image stored somewhere in the brain like a photograph or a slide. It is not even a coded version of such a visual image, the way such an image may be stored on an optical disk, or patterns of on/off switches, to be retrieved by a computer looking up the correct address. Rather the brain remembers the “locomotive” as ephemeral patterns of associations and experiences processed and integrated with other, related patterns all across the brain.
When a “locomotive” is first seen by an individual – child, or adult – a whole series of stimuli are presented to the brain. The brain experiences a “locomotive”. First, there is its shape. Second, there is its color. Those two features alone are sufficient to present a unique pattern as an object. But then it moves, creating brain patterns unique to objects which move. Further, it makes puffs of smoke and clouds of steam. And the noise! Frightening! This activates the limbic system. Patterns experienced while we are frightened (or angry, or in love) are greatly strengthened, therefore more strongly imprinted on our memory.
If sufficiently intense, the pattern may persist (be remembered) for a lifetime. However, such a pattern will need to be reinforced because all learned patterns (in contrast to genetic ones) decay with time.
The brain of an adult (or older child) would already contain patterns abstracted, and quite distinct from, direct experiential patterns such as seeing a locomotive. There would, for example, exist a pattern of abstractions to cover the category “transportation equipment”. This would include all objects which themselves move and in addition can carry or transport other things. A second category of abstractions would include “materials”, causing the visual inputs to look for patterns like bright shiny objects, characteristic of certain metals. Other characteristics might cause the brain to infer that the locomotive was made of iron. A third category of abstractions would consist of a huge constellation of patterns to cover the category “words”, and the sub-category “nouns”, that is words which refer to a person, place or thing. In searching for a word pattern to cover the new experience (the locomotive) the brain might well come up with a new word combination, based on previously stored patterns – for example, “iron horse” as did the North American Indians, and, for that matter, as did the French (Chemin-de-fer).
To differentiate a horse from a locomotive is easy, especially when other sensual experiences are included: A horse smells differently from a locomotive; it feels different to the touch; when put to the mouth – as babies and young children exploring their world would – it would create further unique patterns; it looks different; it moves differently; etc. Nevertheless the idea of creating a new word category “iron horse” to describe a previously unexperienced object such as a “locomotive” is a basic process which we appear to share – judging from the work of the Premacks – with our closest primate cousins the chimpanzees.
The strength of the neural connections making up any given pattern is a function of at least two processes. The first involves repetition: If the same sets of networks are repeatedly activated, they automatically become more efficient and the connections become strengthened. The second involves those experiences which are so important to the organism that they are imprinted virtually forever (or at least until injury or death causes the brain to malfunction) in response to a single event. These involve experiences associated with great danger or great joy. The events leading up to the situation, the scene, the surroundings all make up a great part of the fabric of such an experience and are remembered. It has been established that an increased involvement of the limbic system at the time of an incident – presumably mediated by changes in hormonal states – causes the accident to be remembered better. The juxtaposition of objects, surroundings and events leading up to, for example, a dangerous confrontation generating fear not only create patterns in the brain, but are remembered by the individual as patterns of association. Such patterns constitute a form of wisdom: They allow an individual to “smell trouble ahead.”
The bulk of human intelligence does not derive from the human brain’s ability to deduce conclusions by the application of logic, but rather by its ability to make inferences by perceiving patterns of association. This summarizes the cul-de-sac classical artificial intelligence got into: The primary emphasis was to emulate human thought by working with classical von Neumann type computers, i.e., binary-based logic machines. In contrast, the human brain is an analogue device with an incredibly complex circuitry attuned to picking up patterns of association.
3. Patterns of connections, if they are not periodically reinforced, decay. Reinforcement can come from external stimuli, or internal stimuli – simply thinking about something or remembering, or trying to remember something. Probably sleep is an important factor in helping the decay of trivial connections. Remembering where you parked the car yesterday, or what you ate for breakfast last week, is not likely to make much difference to your future (unless you were poisoned at breakfast, or found a new, secret place to park which would stand you in good stead in the future).
Edelman introduces the concept of neural competition: Some patterns of connection are nourished by stimuli and grow stronger at the expense of other patterns. The weaker ones, which decay, are overridden. Keep in mind that any given nerve or nerve cluster may participate in any one of thousands of patterns, just as a letter or a word may participate in any one of thousands of sentences. It is the weaker connections which disappear, not the neurons themselves.
The “survival of the fittest networks”, i.e., the networks selected for by the brain’s activity, is referred to by Edelman as “neural Darwinism.”
If the above explanations of how the human brain works prove to be correct then one may not only marvel at the simplicity of such an ingenious device – its adaptability and efficiency for learning to recognize complex patterns of association – but also recognize its weaknesses.
First, a neural network which generates memories by strengthening patterns of connections based on constellations of external sensory inputs plus internal associated interpretations is very prone to superstition. Let us consider a hypothetical case. You are driving down a lovely country lane past a farmyard. It is a partly cloudy day. In the distance, to your left, you can hear the faint fumble of thunder, nearby a cock crows, and you hear a tractor sputtering somewhere in the farmyard. The wind rustles through the trees, you see white sheep in the meadow, brown chickens in the farmyard, and a black cat crossing the road ahead. You smell the pungent smell of the farmyard and you admire the thick hawthorn hedges. You come around the bend and – crash into the tractor which has just come in from a side road! Fortunately, you were going slowly and nobody got hurt. But you are shaken. Your adrenalin levels are up and your heart is beating overtime.
In recounting the event later you are struck by the way your mind goes over and over your impressions of that scene just prior to the crash. Nature meant it to be that way. The physiological stress associated with the accident made certain that the lightly weighted connection – the “commonplace” sensory inputs describing the scene above – became heavily weighted. Somewhere in all these “commonplace” sensory inputs lie hidden environmental clues to an impending disaster. The neural network does not discriminate specifics. If in the near future you came around a similar bend, the smell of a farmyard, the crow of a cock, any of these would trigger a state of alert. If a cock crowed just before you had another accident, the neural network would strengthen still further the association between cocks crowing and accidents.
The neural connections may, of course, be strengthened as a result of other processes. To the ancient Greeks, thunder to the left meant a warning from the gods. Our own western culture believes crossing the path of a black cat brings bad luck. Analyzing the accident after the event, depending on your cultural background, you might decide, for different reasons, that the most important environmental clue to the impending disaster – the clue to watch out for under similar circumstances next time – was the thunder, the black cat or the sputtering sound of a tractor. This thinking about the accident, going over it, over and over again, leads to a further strengthening of the neural connections involved. Thus the original associations undergo a process of neural Darwinism. Those pre-accident inputs which appear highly relevant to you are upgraded – if you are one kind of person, you’ll focus on the sputtering tractor engine, if another, on the black cat crossing the road. This subsequent upgrading of certain associations will be at the expense of the others. Thus those pre-accident inputs which appear irrelevant to you will be downgraded and their neural pathways will decay faster.
People who have suffered a trauma, either an injury (physical or psychological) to themselves or to a friend or loved one – perhaps leading to the death of the individual – tend to go over the events leading to the trauma, over and over again. This is an adaptive response – processing the information by running over the circuits, over and over, to extract vital clues from the pattern, clues which might have allowed the trauma to be averted in the first place. Such a process may, however, lead to a pathology by putting a neural network into a self-stimulating resonant cycle – i.e., “thinking in a rut”. One of the antidotes to this phenomenon is sleep. On the other hand, if the process becomes sufficiently intense to constitute a psychopathology, more drastic measures such as drug treatments, or even electroconvulsive shock therapy, may be needed in order to break such a cycle. However, it must be recognized that going over a particular situation repeatedly is a natural form of information processing for a neural network analyzing patterns of connections.
Not unrelated to this is the phenomenon of absent-mindedness, and the tendency for “Freudian slips”. In both instances, the newer, more appropriate, response required of the moment becomes overridden by the more deeply ingrained habits of the mind.
Lastly, an obvious weakness of neural networks is that they are fuzzy and imprecise. All inputs are filtered through numerous nodes and pathways, all of which are likely to have been used for some other purpose previously, or worse, used for similar purposes which may cause a complete misinterpretation of the new inputs. One is reminded of the old tale about Julius Caesar deciding to pardon a soldier condemned to death, by issuing the order: “Execute not, spare!” which unfortunately was garbled in transit to “Execute, not spare!” by moving the comma. A comparable situation, though less critical, must occur over and over again inside our brain.