The previous article summarised the case for claiming that computers were incapable of modelling or simulating the human brain.  (See: http://www.somewhere-else-magazine.co.uk/articles/why_cant_a_golem/)


The point is strongly urged by Susan Greenfield, and is best put in her own words:


“It is this chemical-specific feature of brain function that, in my view at least, makes the brain particularly daunting for those who attempt to model it with computers.  …Different chemicals will be released from different inputs converging on a single cell and active at any moment. In addition to the degree of activity of these inputs, different amounts of transmitter will be released.  Finally each transmitter will dock into its own receptor that has its own characteristic way of influencing the voltage of the target cell.  Thus at every stage there is room for an enormous flexibility and versatility in the brain, using different combinations of transmitter chemicals.”


“This molecular symphony can hardly be regarded as comparable to the scenario inside a computer…[T]he brain is fundamentally a chemical system – even the electricity it generates comes from chemicals.  More significantly, beyond the fluxes of ions into and out of the neuron, a wealth of chemical reactions are occurring incessantly…inside the cell…Second, the chemical composition of the neurons themselves is changing, and hence there is no separate and unchanging hardware, in contrast to a programmable range of software.  Moreover,…computers can ‘learn’ but few are changing all the time to give novel responses to the same commands.”
(Susan Greenfield, The Human Brain: A Guided Tour. Weidenfeld & Nicholson, 1997, pp. 81-82.)


The case seems unanswerable: computers obviously cannot mimic, model or simulate the human brain.  And yet, the very strength of the argument suggests a scintilla of doubt.  If computers are so obviously unable to model the brain, how is it that so many people believe that they can? 


Professor Greenfield’s case is a combination of separate assertions, which look individually cogent, and collectively insuperable.  But the rather dense way in which they are presented suggests that it might be worth taking the trouble to unpack them.  And when this is done, they begin to appear less compelling.


Professor Greenfield’s case can be summarised as follows:


(i) The brain is a chemical system, and therefore does not resemble the processes inside a computer, which are not chemical in nature.


(ii) The electrical impulses in the brain are generated from chemicals in the brain itself, and this is not the case in a computer.


(iii) The various inputs into a given neuron in the brain involve different chemicals, in different amounts, affecting the target cell in different ways depending on the specific receptors involved.


(iv) The many chemical reactions incessantly taking place inside a neuron mean that the neuron itself is constantly changing, whereas the hardware inside a computer is essentially permanent.


(v) Only a few computers are capable of changing all the time to give novel responses to the same commands, whereas this is typical of neurons.  Therefore, most computers cannot model this aspect of neuron function. 


Looking first at (i) and (ii), as statements of fact these are indisputable.  However, as arguments against the possibility of simulating brain activity in a computer, they are less convincing.  The implicit logic behind them runs something like this:


The brain is a chemical system which generates its electrical energy from chemicals (for short: the brain has property C).


Anything which simulates something with property C must itself have property C.


No computer has property C.


Therefore, no computer can simulate the brain.


But if this logic is correct - and it appears unimpeachable - it should be applicable to other situations.  So let us apply it to the situation in an oil refinery:


An oil refinery is a complex chemical system.


Anything which simulates a complex chemical system must itself be a complex chemical system.


No computer is a complex chemical system.


Therefore, no computer can simulate an oil refinery.


Now the conclusion to this argument is clearly false: computers do simulate, and indeed to some extent control, oil refineries.  So where is the error? 


There does not seem to be anything wrong with the first and third premiss.  So the second one, “Anything which simulates something with property C must itself have property C”, must be at fault.  Indeed, on consideration, there is no reason why the chemical nature of the brain should be a bar to a computer’s modelling it; we would not require that a computer be run by steam in order that it be capable of simulating the processes occurring inside a steam engine, and it is surely naïve to require that a computer be composed of a chemical soup, or of biological cells, in order that it be capable of simulating the behaviour of that soup, or those cells.


Having disposed of (i) and (ii), let us examine (iii): the various inputs into a given neuron in the brain involve different chemicals, in different amounts, affecting the target cell in different ways depending on the specific receptors involved.


Greenfield’s argument appears to be that a computer cannot simulate a cell (or cells) in the brain, because a computer’s internal signals are conveyed by just one electrical signal, whereas the cells in a brain communciate using many different chemical messengers, delivered in variable quantities, and with variable effects depending on the nature of the target cell. 


But the oil refinery analogy is also applicable here.  We may as well argue that because there are various chemicals arriving at different destinations in differing quantities in a refinery, with differing effects depending on which reaction vessel they are delivered to, a computer cannot simulate their effects; but this is known to be untrue. 


The resolution of the paradox is that a computer does not need to imitate the action of the brain, in the naïve sense of using the same mechanisms for communication between its components as a brain uses for communication between its cells, in order to simulate it.  A computer is capable in principle of modelling a neuron with many different inputs; the chemicals, and their effects, simply need to be appropriately coded.  The actual calculations may be complex, and even impracticable beyond a certain degree of complexity, which may in terms of today’s technology fall far short of the complexity of the human brain.  But that is a different argument.


What about (iv)?  This states that “the many chemical reactions incessantly taking place inside a neuron mean that the neuron itself is constantly changing”. 


Presumably Greenfield’s point here is that because a neuron is constantly changing, it cannot be simulated by a computer, whose components (ie hardware) do not change at all, at least unless an operator physically removes one module and adds another. 


Again, this argument, if valid, would show that a computer cannot model a chemical reaction vessel, whose contents are constantly changing in consistency, temperature and quantity.  But we know that to be untrue.  The answer is that a computer with fixed hardware can model a cell, or system, with changing “hardware”; it does so by simulating the system in its memory, which is capable of storing the characteristics of virtual neurons with an indefinite degree of flexibility.


The answer to (v) is similar.  Computers are capable of changing their internal states over time, without external commands; it is possible to programme them with the software to do this on the basis of the internal clocks which all computers possess, moderated by whatever mathematical modelling the programmer wishes to build into their memories.


Professor Greenfield has a slightly different argument against the computer analogy elsewhere in “The Human Brain” (pages 110-111):


“Even when grown in a dish - in tissue culture - brain cells will send out their axons.  Therefore, using a time-lapse video, it is possible to observe directly brain cells reaching out to their neighbours to establish contact.  Viewed in this way, it is hard not to anthropomorphise the developing cells.  As they move on film, they appear highly purposeful, yet with the fragility of spun sugar, moving across the screen with alarming speed while literally feeling their way by means of fluted, weblike endings that undulate and flutter as they make their inexorable progress. … When looking at such films, it is hard to understand how anyone could view the brain as a computer or even compare it with one”.


It is easy to be carried away by the poetry of Greenfield’s description into an uncritical acceptance of her final remark.  But it is worth suspending our delight and wonder at these undoubted marvels of brain biology for long enough to question whether the last sentence is actually entailed by what has gone before. 


The growth and development of the human eye could be painted in colours as florid and exciting as those Greenfield uses for the growth of neurons: the more legitimately, given that the retina is in fact an organic outgrowth of the brain.  Suppose that someone were to observe the “purposeful” growth of retinal cells in a petri dish and conclude that “when looking at such films, it is hard to understand how anyone could view the eye as a camera or even compare it with one”.  Put in these terms, the fallacy becomes obvious: the development of the eye has little relevance to its optical functions when fully formed.  The eye can be compared with a camera, and is so compared by optometrists with considerable practical success; in some ways the eye behaves very like a camera, though of course in other ways it does not.  The point is that it is fruitful, and helpful to our understanding of how the eye functions, to analyse it as an optical system.  The point of comparing the brain to a computer is to test whether or not the comparison is fruitful in understanding the functionality of the brain, a point to which Greenfield’s argument here is irrelevant. 


It must be acknowledged that computational approaches to the brain have considerable weaknesses.  Neural net models of the brain have had some success in simulating learning processes; they appear to be consistent with Karl Lashley’s findings on the graceful degradation of memory when the cortex is damaged.  But these models are very crude approximations to what goes on in any brain, let alone the human brain, both because the number of “neurons” is relatively tiny, and because current models lack the complexity described in Greenfield’s points (iii) and (iv) above.  The models could in theory incorporate some of these features; but as yet we know too little about how the brain actually works to incorporate them accurately.  The gap is due to our ignorance, not the fundamental limitations of the computer simulation process; this is the real weakness in current approaches.


A further, more basic gap becomes apparent when we compare the brain with comparably complex systems where we do have accurate mathematical models of their function.  Statistical thermodynamics enables physicists to predict the properties of gases, and some solids, with great accuracy, even though their behaviour involves the interactions of trillions of atoms and molecules occurring millions of times every second.  This is possible because physicists realised 150 years ago that it was not necessary to follow the individual movements of molecules in order to predict the behaviour of the bulk substance, but merely to calculate certain macroscopic quantities of a statistical nature such as temperature, pressure and entropy.  It is attractive to speculate that similar unifying concepts may one day enable us to attain a qualititively better understanding of the brain, and to dispense with the puny neural models in use today, which are both too literal (in trying to imitate the micro-structure of the brain) and not literal enough (in failing to imitate its full complexity).


I suspect that we are in a similar position to those nineteenth century writers who speculated on the possibility of manned flight.  There were many at the time who criticised the fumbling attempts of the pioneers to achieve powered flight, as demonstrating only its impossibility.  The Astronomer Royal famously “proved” that man could not fly, a decade before the Wright brothers showed him to be mistaken.  Somebody once said that when scientists claim that a thing is possible, they are generally right, but when they pronounce something to be impossible, they are always wrong.  If history is any guide, we should hesitate a very long time before accepting the dictum of even such an eminent authority as Professor Greenfield, concerning the potential of computers to enhance our understanding of brain function.