The existence of a selective impairment that affects only certain kinds of association and not others, or only certain kinds of neural encoding (for sounds) but not others, suggests that the neural wiring that would normally serve those associations or encodings is not quite right. Perhaps the wiring has been laid down in a way that does not readily support those kinds of associative processes. Because the wiring is, in part, specified genetically, the problem should have a genetic basis, and it does. If you have a dyslexic parent, there is a high chance that you will be dyslexic too. One fact about dyslexia that is clear is that it is heritable.
But dyslexia is not the only language problem in childhood that is heritable. Just as acquired language disorders can affect either spoken language or written language, or both, so in childhood, there are disorders which are not confined to the written language.
Language-specific impairments in childhood
There is a syndrome commonly referred to as specific language impairment (SLI), which is not confined, like dyslexia, just to reading, although it is often accompanied by a deficit in reading. A child will be classified as having SLI if there are no other obvious deficits (they have normal vision, normal hearing, a normal IQ in tests that do not require verbal skills, and so on). It is unclear, though, whether SLI should really be thought of as a syndrome or, instead, as an umbrella term that is applied to any deficit that affects language use. Numerous studies have attempted to establish the cause of the deficit and, unsurprisingly, many have been found. Some SLI children have difficulty discriminating between rapidly changing speech sounds, although standard hearing tests would show their hearing to be entirely normal. Other children seem to have a problem with grammatical function words and inflections, similar to some of the acquired aphasias described earlier. Their speech is slow and laborious, and their comprehension is often poor. Yet other children fall somewhere in between.
In the mid-1980s, Myrna Gopnik at McGill University in Montreal reported an entire family, spanning several generations, with a high incidence of SLI amongst its members. The pattern of incidence suggested that SLI had been transmitted genetically from one generation of the family to the next. This finding, coupled with the claim that the family had the version of SLI which affected use of grammar, led to all sorts of speculations about the existence of a `grammar gene'. There are two reasons to be cautious about such speculation. First, it seems that not all the individuals in this family suffered from the same brand of SLI. But, more importantly, perhaps there is no more a grammar gene, or a language gene come to that, than there is a reading gene. After all, dyslexia is heritable, but reading has only been around a few thousand years, so it is inconceivable that evolution has had time to encode anything specific to reading within our genetic make-up. So whatever it is that determines genetically whether or not we are dyslexic is not reading-specific. Similarly, whatever it is that determines genetically whether or not we are likely to suffer some other language-specific impairment may not even be language-specific. It may just be specific to the kinds of wiring that happen to be well suited to the associative process that underlie our use of language.
Other childhood impairments
A very different problem that can arise in childhood is a stutter. Ordinarily, a stutter is considered a speech impediment. There is no equivalent disorder involving any other kind of muscular movementit is articulation-specific. But it can also be specific to a particular language, and in at least one case, a stammerer stuttered badly when speaking one language, but not when speaking another. Many adults who stutter badly during normal conversation can sing, or speak in a `funny voice' without stuttering at all (the author P. G. Wodehouse wrote a short story based on such a case).
There are different kinds of stutter, and they probably each have a different cause. Some theories have proposed that a stutter arises when the right hemisphere has taken more control than usual over the articulators (the lips, tongue, and so on). Other theories propose that it has to do with faulty feedback between the signal that initiates the articulation, and the signal received when the articulation happens (if normal adults are asked to speak aloud whilst hearing their own voice played back over headphones after a slight delay, they too stutter). Most theories agree that it has something to do with mis-timing during the articulation process. The fact that only a small proportion of stutterers are female (around 20%) suggests, once again, a genetic component.
Another impairment that can affect articulation is not languagespecific at all-children with a hearing impairment often have impaired articulation. When learning to speak, it is inevitable that, if what you hear and attempt to imitate is distorted, what you speak will be distorted too. Possibly the most common kind of hearing impairment in children is temporary. Children with otherwise normal hearing, but who are prone to ear infections, may suffer from glue ear. Here, the middle ear (the part on the inside of the ear-drum) fills up with a sticky fluid that prevents the normal transmission of the vibration of the ear-drum to the inner ear, where nerve fibres respond to the vibration. Often, glue ear is first noticed because the child has an obvious hearing impairment. But sometimes it is noticed because the child's speech is not as clear as would be expected. Often, there are considerable knock-on effects in school. Fortunately, glue ear can be treated relatively easily, and most children catch up quite quickly once the problem has been eliminated.
Sadly, not all impairments to language are so easily treated.
Almost anything that one can imagine going wrong with language can go wrong. And some things one would not imagine could go wrong do. The range of impairments is vast, and often, especially following brain injury, language is not the only faculty that has been impaired. It is all very well to describe, clinically and unemotionally, the nature of these impairments, but imagine waking up one day and discovering that it had happened to you. Or imagine growing up and losing out at school because you could not take notes easily, or could not understand the teacher, or could not express yourself adequately. Imagine your child having to grow up that way. None of us who are unimpaired can imagine what it is like to hear people speaking but be unable to understand them, or to know what we want to say but be unable to say it. For most of us, the ascent of our own personal Babel is unimpeded, and we should be thankful for that.
Wiring-up a brain
The average adult human brain weighs around 1.3 kg, and contains 10 billion or so nerve cells. Each nerve cell, or neuron, can connect to, and so stimulate, anything between a few hundred and perhaps 100 000 other nerve cells. And each neuron can itself receive connections from up to that same number again. Extend this to all 10 billion cells, and it is surprising that anything as vast and complex could work at all. But evidently it does. Because there is little else in the brain apart from neurons, we have no choice but to accept that they, and the manner in which they interconnect, are responsible for the mental feats of which we are capable. But there is rothing particularly special about neurons: if you stimulate one enough, it will stimulate the other neurons to which it is connected. Yet from this come our mental faculties. There is one further property of the brain that is crucial-even in an adult brain the wiring between the neurons is constantly changing. If it could not change, we could never learn.
We are not born knowing the language we shall end up using. We learn that language. Just as we learn about the world within which we shall use it. The meanings which we evoke with the words of our language are simply patterns of neural activity. These patterns reflect the accumulated experience of the contexts in which those words are used, and as such they have gradually changed with those experiences (see Chapter 9 for further details). But for patterns of neural activity to change, the patterns of neural connectivity and neural transmission that underlie those patterns of activity must also change. Ultimately, it is these changes that allow us to learn from experience. And learning from experience underlies just about everything we do. So how do these changes come about?
&nbs
p; Neurons are a little like sunflowers. The flower corresponds to the body of the neuron. The stem is the main length of the neuron down which neural impulses are transmitted to a mass of roots. These connect to other neurons (via the equivalent of the sunflowers' `petals') to which those impulses are transmitted.
Despite the complexities of the neurochemical processes that underlie neural transmission, there are just three principles at work. The first, and most obvious, is that neurons send impulses to the other neurons to which they are connected. The rate at which impulses are sent corresponds, in a sense, to the `strength' of the signal. The second is that an impulse from one neuron can either make it more likely that another neuron will generate an impulse of its own, or less likely. Which of these it is depends on the kind of connection (it can be an excitatory connection, or an inhibitory one). The third principle is perhaps the most important. The connections can change in response to the surrounding neural activity-new ones can grow, especially in the first two to three years of life, although it happens in adulthood too; existing ones can die back (again, this is probably more common in younger brains); and the sensitivity of each connection can change, so that a neuron will need to receive either a stronger or a weaker signal across that connection before it generates its own impulse.
But from these three principles, are we any closer to understanding how a brain can wire itself up for language (or for anything else, come to that)? This is where artificial brains come in.
Inside an artificial neural network
Artificial neural networks exhibit the same three principles introduced in the preceding section. But the neurons are quite different. For a start, the most common neural networks (the `artificial' will be omitted from now on) are simulated. A computer program keeps track of which neurons there ought to be, what each should currently be doing, which should be connected to which, and so on. These simulated neurons are a lot simpler than their real counterparts. The purpose of these neural networks. is not to simulate the precise workings of the brain. What matters is that a signal is passed from one neuron to each of the others it connects to, or the fact that the sensitivity of a connection can change. There is no attempt to model the process by which it changes.
Artificial neurons are very much simpler than real ones. The computer works out how much stimulation a neuron should receive, and allocates that neuron a number to reflect how active it should be. Real neurons are similar-the rate at which they generate neural impulses, the strength of the signal, changes as a function of how much stimulation they receive. This number, the neuron's activation, is calculated on the basis of how active each of the neurons feeding into it is, and how `strong' each of these connections is. The connection strength reflects the fact that the sensitivity of real neural connections can vary. It is a little like the volume knob on an amplifier. The higher the volume, the louder the signal. But the difference is that in these artificial networks, the connection strength is just a number. If it is positive, it is an excitatory connection. If it is negative, it is an inhibitory connection. If it is zero, then that is the same as if there were no connection between those two neurons.
That, briefly, is the underlying physiology of a neural network. The fact that a computer keeps track of what is going on within an artificial neural network, and does all the working out, is immaterial. We could instead build an artificial neural system which consisted of physically distinct artificial neurons, with complex interconnections that enabled the neurons to, in effect, add, subtract, and multiply (equivalent interconnections exist within the brain). It is just much easier to have a computer simulate all this. But the important point is that the computer is simply doing what a real network could in principle do for itself.
How networks work
The easiest way to figure out how neural networks work is to go through an example. There are many different kinds, and we shall take as our example one that could be used, for instance, to learn letter-tosound correspondences. Typically, the neurons in these networks are separated into distinct groups. In our example we shall use just three such groups. One is going to act as the `eyes' of the network-patterns of activity across the neurons in that group will represent the letter that the network is `seeing'. Another group is going to represent the phonemes that the network is supposed to output (perhaps to a speech synthesizer). The third group will be intermediaries between the letter neurons and the phoneme neurons.
In real brains, there would probably be thousands of neurons in the chain, but the advantage of artificial neural networks is that they can be very much simpler than real brains. The only route from the letter neurons to the phoneme neurons in this example network is via the intermediary neurons. In principle, we could allow a more direct route, but we shall not do so here. Another thing we shall not allow in this example are direct connections from one letter neuron to another, or between the phoneme neurons.
That is the basic anatomy of the network. How does it work? We can start off by assuming that it has not yet learned anything. We must also assume that when the network `sees' a letter, the computer activates the letter neurons, giving each neuron a particular amount of activation. Each letter of the alphabet would have its own unique activation pattern. We shall return shortly to why the computer allocates one pattern, rather than another, to any one letter.
We can now work through what happens when the network sees an L. First, the letter neurons will each be activated, by different amounts, according to the pattern of activation that has been allocated to that letter. Next, the computer will look at each neuron in the intermediary set of neurons, work out how much stimulation each one is receiving from all the letter neurons that connect to it (taking into account each connection's strength), and activate it accordingly. For each neuron, it takes the activation value of all the neurons connecting to it, multiplies each of those values by the appropriate connection strength, adds the results of all these multiplications, enters the grand total into an equation which converts it to a number between 0 and 1, and sets the neuron's activation to this final value. Once it has done this for all the intermediary neurons, it does the same thing again for the phoneme neurons connected to the intermediary neurons. In this way, the pattern of activation across the letter neurons spreads, via the intermediary neurons, to the phoneme neurons.
The final pattern that develops across the phoneme neurons will mean absolutely nothing. The network has not learnt anything yet, and the connection strengths are all just random. So the activation pattern across the phoneme neurons would also be random. But if the network had learnt what it was intended to learn (and we shall come to how it would do this shortly), the pattern across the phoneme neurons when the network was seeing the letter L would have been a pattern that was supposed to correspond to the phoneme /1/. It would be a pattern that the computer had previously allocated to that phoneme, in much the same way that it had allocated one pattern to the letter L, another to M, and so on. The learning process would have taken a random set of connection strengths, and would have managed to change them so that a particular pattern across the letter neurons (the pattern for L) would spread through the network and cause a particular pattern across the phoneme neurons (the pattern for /1/).
There is an important consequence of this last fact. If the connection strengths start off as random, they will scramble up the pattern allocated to L when they transmit it to the intermediary neurons. Whether it is one pattern or another makes absolutely no difference--it will still be scrambled. But if the network can learn to change the connection strengths so that it activates a specific pattern across the phoneme neurons in response to a specific pattern across the letter neurons, even when the connection strengths started off as random, it would not matter what pattern was initially allocated to any individual letter, as long as it was different from the pattern allocated to any other letter. This is just as well when it comes to thinking about our own brains and the activation patterns that they start out with. As long a
s our senses are consistent, it does not matter what patterns of neural activity they cause, so long as the same sensation gives rise to the same pattern of activity, and there is some way of changing the neural connectivity from its initial (potentially random) state to its final (most definitely non-random) one. Exactly how this happens in the artificial case is described next.
How (some) networks learn
The way in which networks like the one in our example learn is surprisingly simple. But before seeing how they learn it is as well to consider what they learn. The sensitivity of each connection within the network determines the precise pattern of activation that forms across the phoneme neurons in response to a particular pattern of activation across the letter neurons. So if you want the network to produce a particular phoneme pattern in response to a particular letter pattern, you have to get the connections just right. And that is what networks learn to do. They can learn to set the sensitivities of their own connections so that, eventually, the network can encode many different letterphoneme pairings using the same set of neural connections.
The Ascent of Babel: An Exploration of Language, Mind, and Understanding Page 25