Oxford university press how the br ain evolved language

Figure 11.1. 
Fragment of a semantic network. 
Quillian’s one representation scheme greatly simplified the organization 
of a computational lexicon. It could encompass verbs, nouns, and other parts 
of speech. Moreover, the semantic network was also a compact representation. 
Instead of repetitively coding tens, and possibly hundreds, of features on each 
word in the lexicon, features could be inherited along ISA links. Thus, each 
meaning element needed to be represented only once in the net. For example, 
having the links dog ISA mammal and mammal ISA animal, we don’t need to 
represent the fact that a dog can follow links from mammal to the further infer­
ence that a dog ISA animal. 
In psychology, Quillian’s work was closely followed by a series of “spreading 
activation models” of word association. Morton (1969, 1979) proposed his 
“logogen model” as a psychological account of word recognition during read­
ing. In this model, concept nodes (logogens) fired when they detected a word, 
and the more they fired, the more easily they would fire the next time. In this 
manner, they learned the word they detected. Contextual activation was also seen 
as spreading between logogens along semantic links. Such spreading activation 
models featured a Hebbian explanation of why frequent words are recognized 
more quickly than infrequent words, and they also offered an explanation of 
certain “priming effects.” For example, subjects recognize the word nurse more 

TRUTH  AND  CONSEQUENCES 
•  169 
quickly after having first been “primed” by hearing or seeing the word doctor 
(Swinney 1982; also see Small et al. 1988 for a review and related papers). 
Perceptrons 
Ever since Ramón y Cajal, it had been obvious that the brain is a massively 
parallel processor, but no one knew what this meant. There had been a few early 
attempts to model parallel computation (e.g., McCulloch and Pitts 1943), but 
it wasn’t until the late 1950s that researchers began to think seriously of the 
serial, digital computer as a starting point for parallel models of the brain. In 
1958, Rosenblatt popularized the most famous early parallel-computing model, 
the perceptron (Rosenblatt 1958). For a brief time, parallel computer technol­
ogy competed with serial computer technology, and as early as 1965, the Illiac 
IV, a 64,000-element parallel computer, became operational. 
It seems, however, that no one could figure out how to program the Illiac 
IV, and in 1969, Marvin Minsky and Seymour Papert of MIT published Percep-
trons, certifying what experience had found meaningful to also be True. Find­
ing that perceptrons could not even compute the elemental and essential 
Boolean operation of XOR, Minsky and Papert concluded that “there is little 
chance of much good coming from giving a high-order problem to a quasi-
universal perceptron whose partial functions have not been chosen with any 
particular task in mind” (1969, 34). As it happened, only three years later a 
celebrated young associate professor in Papert’s laboratory published the first 
of several papers which showed not only that the brain was a parallel proces­
sor but also that it computed XORs ubiquitously (Grossberg 1972b, 1972c). 
But no one was buying parallel computing any longer. After Minsky and 
Papert’s critique, funding for parallel computation research dried up, and fur­
ther progress in the field was not widely recognized until 1982 when Hopfield 
published his theory of “content-addressable” memory. Hopfield’s memory was 
reminiscent of earlier work in perceptrons, but in the intervening fifteen years 
microprocessors had been developed. IBM had just introduced its personal computer, 
and it was now easy to imagine bringing massively parallel computers to market. 
Other researchers, following these models, extended the basic, single-layer 
perceptron into multilayer perceptrons. In an unpublished doctoral dissertation, 
Hinton (1977) had described such a system, and in the new climate, Ackley, 
Hinton, and Sejnowski (1985) elaborated that system as an extension of the 
Hopfield model. By sandwiching a layer of Hopfield-like neurons between lay­
ers of input and output neurons, the resulting multilayered perceptron, or 
Boltzmann machine, could not only remember but also actively classify. At nearly 
the same time, a number of researchers, most prominently Rumelhart, Hinton, 
and Williams (1986), added to this architecture an error correction technique 
known as back-propagation. Back-propagation models employ a fundamentally 
cerebellar design and so are not particularly useful for modeling higher cogni­
tive functions like language (Loritz 1991). Nevertheless, they are relatively easy 
to implement on serial computers. As a result, they have found a ready resonance 

170  • 
HOW  THE  BRAIN  EVOLVED  LANGUAGE 
among computer scientists, and they furnished psychologists and linguists (e.g., 
Goldsmith 1993, 1994) with new metaphors of mind and language. In the sand­
wiched, or “hidden,” layers of multilevel, back-propagating perceptrons, the 
convergence of inputs and divergence of outputs are so complex that simple and 
direct interpretation of their activation patterns becomes impossible: the asso­
ciations between input and output are effectively hidden from the researcher. 
10
Which brings us, by a commodius vicus of recirculation, back to meaning. It is 
much the same in real thought. Between the input and output layers of language, 
the intricate connections of our meaning are hidden, shrouded in complexity. 
Even if I could communicate, instant by instant, every synapse of meaning, every 
past and possible x
i
 and z
ij
 in my brain, my meaning would still be hidden in the 
7,111,111
 different ways you could interpret those weights. Of course, when our 
tangled nets of words and concepts happen to reach a place of relative under­
standing, we can give that place a name. And when we look back on the places 
we have been and the path we have followed, we can call it logic. If we keep our 
names simple enough, if we only seek to agree on what is 1 and what is 0, then 
we might even find that this path of logic leads to a kind of Truth. But Truth 
does not guarantee survival, and this linear logic is not Meaning; it is only a trail 
of linguistic cairns. 
This brief history of linguistic science gives us pause to reflect on how thought 
is like meaning. In the preceding chapters of this book, I have described thought 
as a social state of neurons, neurons involved in acts of communication. In this 
chapter, I have described meaningful linguistic theories as a social state of Homo 
loquens. Thought and meaning are both only stable when neurons or people are 
in a state of communicative resonance. This standing wave of communication 
may appear to be an eternal, true state, especially to those individuals who are 
locked in its resonance, but at the periphery, the environment is always chang­
ing, and there are other inputs which other populations are detecting and en­
coding. All it takes is an unexpected turn of events—a burst of nonspecific arousal, 
a Columbus sailing west and arriving east—for a revolution to occur. 
We shouldn’t be surprised to find these similarities between thought and 
meaning, between neural society and human society. After all, a human being 
is just a neuron’s way of making another neuron. This may be self-similarity 
pushed too far (I am reminded of Daniel Dennett’s “a scholar is a library’s way 
of making another library”), but this is the direction in which self-similarity 
leads. Even it doesn’t form a useful basis for either neurosurgery or public 
policy, it does often supply a useful metaphor, one we can use to develop mean­
ing, if not truth. 
I am often asked by my students, “Is it true we use only 10% of our brains?” 
To which I respond, “Does a society use only 10% of its people?” Of course, 
some do, but in the main, most societies use almost all of their people almost 
all of the time. The real question is how does it use them? Lateral competition 
exists in sibling rivalry and office politics as surely as it exists in the cerebrum, 
and the human organism swims ahead in history, oscillating its tale from left 
to right, unsure of where it’s going but trying always to survive. 

WHAT IF LANGUAGE IS LEARNED BY BRAIN CELLS
?
•
171
• 
T 
W 
E 
L 
V 
E
• 
What If Language
Is Learned by Brain Cells?
The generative deduction held that children learn their first language effort­
lessly. As the theory went, an innate language acquisition device (LAD) stops 
operating somewhere between the age of six and twenty. Children, thanks to 
their LAD, were supposed to “acquire” language perfectly and “naturally” 
(Krashen 1982). Adults, however, were left with no alternative but to “learn” 
second languages in an “unnatural” and suboptimal manner. Thus, generative 
theory was forced to posit two mechanisms, which it could not explain, much 
less relate: one to account for the facts of child language “acquisition” and 
another to account for the facts of adult language “learning.” 
Adaptive grammar explains both child and adult language learning with 
one set of principles which evolves similarly in both ontogeny and phylogeny 
and operates similarly in both brain and society. After all, it is not clear by what 
yardstick generative philosophers measured child labor. We adults say we find 
second languages harder to learn as our years advance, but it may just be that 
we value our time more as we have less of it. By simple chronological measure, 
normal children do not learn language all that quickly. It takes children six 
years or so to learn the basic sounds, vocabulary, and sentence patterns of their 
mother tongue. By some accounts, this process extends even beyond the age 
of ten (C. Chomsky 1969; Menyuk 1977). Adults can learn all of this much faster 
than children (Tolstoy, it is said, learned Greek at the age of eighty), even if 
there is some debate as to how well they learn it. 
There is a widespread consensus that adults never learn to pronounce a sec­
ond language well—spies and opera singers notwithstanding—but few would 
argue that fastidious pronunciation is a cognitive ability of the first order. Adap­
tive grammar assigns fastidious pronunciation to the cerebellum, and just as 
few violinists who take up the instrument in retirement achieve the technical 
capacity of a five-year-old prodigy, the adult who attempts to emulate Tolstoy 
may find his cerebellum regrettably nonplastic: as we have seen, the cerebel­
171 

172  • 
HOW  THE  BRAIN  EVOLVED  LANGUAGE 
lum lacks the capacity to learn new patterns through rebounds. In adaptive 
grammar’s analysis, the cerebellum is as essential to language as rhythm, and 
in this chapter we will even relate it to several language disorders, much as Kent 
et al. (1979) related it to plosive ataxia in chapter 8. But pound for pound, the 
cat has a bigger cerebellum than Homo loquens, so the cerebellum seems an 
unlikely site for any uniquely human language acquisition device. 
On the other hand, the cerebrum is never so plastic as at the moment of 
birth, when it is—sans sight, sans smell, sans taste, sans dreams, as Changeux 
(1985) puts it—as rasa a tabula as Locke could have imagined. Its ability to adapt 
thought and language to an unstable environment has secured Homo loquens a 
singular history of survival. Not even T. Rex ruled his world with such power 
over life and death. In the preceding chapters, we have tried to follow that 
history through evolutionary time, through the phylogeny of Homo loquens. We 
will now recapitulate those chapters in an examination of the ontogeny of lan­
guage in the individual human. 
Prenatal Language Learning 
There is no disputing the fact that language is innate. Without genes, rocks 
would be the highest life form. But there are only some 10
5
 genes in the entire 
human genome, and only some 10
3
 of these are uniquely human. Even if all of 
these were genes for language, and even if they are allowed to operate in com­
binations, there simply aren’t enough uniquely human genes to specify uniquely 
human language. For language, we must still recruit the genes of our phyloge­
netic ancestors, and over this still-sparse genetic scaffolding, language must still 
be learned. This learning begins even in the womb, as the environment im­
pinges upon brain development. 
The environment of the womb is not all so hostile as the primordial soup. 
Almost from the moment of conception, the mother’s heartbeat envelops the 
child, and its first neurons learn the rhythm of humanity. This is not quite yet 
the rhythm of language, to be sure, but soon the child’s own heart begins to 
beat, and the first rhythmic foundations are laid for language and all serial 
behavior. The human child is conceived, and the fertilized egg begins to di­
vide into multiple cells, forming first a blastula and then a two-layered gastrula. 
Within two weeks, the vertebrate notochord begins to clearly differentiate itself 
from the rest of the gastrula. By four weeks, the cerebrum has differentiated 
itself from the notochord and attained about a size and shape proportionate 
to an Ordovician fish. At the same time, an otocyst forms from the ectodermal 
membrane. This otocyst will become the inner ear. 
By seven weeks, the fetal brain has achieved roughly the cerebral develop­
ment of a reptile. Within another week, the cochlea will be completely formed 
and coiled. By twelve weeks, the cerebrum will have achieved rough propor­
tionality to the brain of an adult cat, and by about four or five months, the child’s 
full complement of neurons will have been produced by mitotic division. 

WHAT  IF  LANGUAGE  IS  LEARNED  BY  BRAIN  CELLS
?  •  173 
Critical Periods 
In the wake of Hubel and Wiesel’s work suggesting a critical period for the orga­
nization of visual cortex, Lenneberg (1967) suggested that a similar critical 
period might exist for language acquisition. It was originally widely believed 
that critical periods were, like growth and language development, essentially 
irreversible. Just as there was assumed to be a detailed genetic plan for the 
development of notochord and otocyst, there was assumed to be a genetic plan 
for the organization of the brain, down to the detail of Hubel and Wiesel’s 
ocular dominance columns. Lenneberg hypothesized that the organization of 
language cortex into yet-to-be-discovered grammar-specific structures resulted 
from a similar genetic plan. Generative philosophy held that language is not 
learned so much as it grows in the child. 
As it turns out, however, the Hubel-Wiesel critical period is reversible. 
Kasamatsu et al. (1979) poisoned the noradrenergic arousal system of young 
cats and observed that this diminished plasticity in accordance with the Hubel-
Wiesel critical period. But then they added noradrenaline to the cortex of these 
adult cats, after their critical period had supposedly ended, and they found 
that this restored plasticity. Coupled with the effects cited in chapter 3, where I 
noted that noradrenaline can increase the signal-to-noise response ratio of 
neurons, adaptive grammar interprets Kasamatsu’s nonspecific noradrenalin 
suffusion as having effected a kind of nonspecific arousal (NSA) and an ensu­
ing plastic rebound. Such a reversal of plasticity would be a most unusual prop­
erty for a growthlike genetic process. But as we have seen, no detailed genetic 
plan is necessary for ocular dominance columns and tonotopic maps to develop. 
They can self-organize with no more detailed a plan than that of an on-center 
off-surround anatomy. 
Lateralization 
Broca’s discovery that language is normally lateralized to the left cerebral hemi­
sphere has attracted curiosity and speculation ever since its publication, and 
the fact of lateralization figured prominently in Lenneberg’s speculations. But 
no sooner had Lenneberg proposed that lateralization occurred during child­
hood than Geschwind and Levitsky identified a lateralized cerebral asymme­
try in the superior planum temporale of human and animal fetuses (Geschwind 
and Levitsky 1968; Geschwind 1972; Geschwind and Galaburda 1987). This area 
of the left hemisphere between Heschl’s gyrus and Wernicke’s area was found 
to be larger than corresponding regions of the right hemisphere in 65% of 
the brains studied (figure 12.1). The left and right areas were found to be equal 
in 24% of the cases, and a reverse asymmetry was found in 11%. Noting that 
this enlarged planum temporale was less reliably found in boys, who are also 
more prone to be left-handed, language-delayed, and dyslexic, Geschwind 
hypothesized (1) that abnormal symmetry (non-lateralization) could be caused 

174  • 
HOW  THE  BRAIN  EVOLVED  LANGUAGE 
Figure 12.1. 
Cerebral asymmetry in the superior planum temporale (PT). 
(Geschwind and Levitsky 1968. Reprinted by permission of the American Associa­
tion for the Advancement of Science.) 
by the same effusion of testosterone that, while effecting sexual development 
in utero, also affected brain development, and (2) that many, if not most, left­
handers might have been brain-damaged in infancy in this way or another. 
At first it was thought that such physiological lateralization of language into 
the left hemisphere was a critical and uniquely linguistic process. But further 
research discovered that birds and bats and all manner of beasts exhibit brain 
asymmetries, and lateralization is no longer widely thought to be a language-
specific process. Rather, adaptive grammar suggests that all serial behaviors are 
normally lateralized to one hemisphere because, at their lowest levels, serial 
long-term memory (LTM) primacy gradients optimally form within a barrel’s 
inhibitory surround. This would encourage (but not strictly require) atomic 
serial patterns like syllables and words to organize in compact, intrahemispheric 
regions like Broca’s area, Wernicke’s area, and the nominal and verbal regions 
of temporal cortex described by Damasio et al. (1996). If Homo loquens needed 
to locate a few extra serial LTM gradients anywhere, the larger superior planum 
temporale, which surrounds the left koniocortex, and the planning regions of 
Homo loquens’ extensive frontal cortex in and anterior to Broca’s area would be 
good places to put them. 
Kim et al. (1997) have found evidence supporting this explanation. They 
conducted an fMRI study of six fluent early bilinguals who learned their second 
language (L2) in infancy and six fluent late bilinguals who learned their L2 in 
early adulthood. MRI scans of the late bilingual subjects speaking both L1 and 
L2 (which varied widely) revealed significantly distinct regions of activation, 
separated by a mean of 7.9 mm, for each language within Broca’s area. Late 
bilinguals’ L1 morphophonemic motor maps had developed in close proxim­
ity, as if their extent were limited by some factor like inhibitory radius. Their 
L2s developed in a separate area, as if the L1 area was already fully connected. 

WHAT  IF  LANGUAGE  IS  LEARNED  BY  BRAIN  CELLS
?  •  175 
Early bilinguals did not exhibit two distinct regions; presumably, their morpho-
phonemic motor maps all developed together in a single area.
1 
Although left lateralization may be a normal event in language develop­
ment, it does not seem to be a necessary event. In deaf sign languages Heschl’s 
gyrus and the planum temporale are only marginally activated. If anything, sign 
language seems to be doubly lateralized—with its temporal, planning functions 
lateralized to the left hemisphere and spatial functions lateralized to the right. 
In childhood aphasia, we also find the heartening fact that recovery is possible 
and often total. The plasticity of the brain allows the child aphasic’s right hemi­
sphere to completely take over the function of the better-endowed but injured 
left hemisphere. Left-handers wind up being, if anything, slightly superior to 
right-handers in their verbal skills. The only necessary loss seems to be a small 
delay in the (re)learning of language, in the translocation of some atomic 
primacy gradients to the contralateral hemisphere. If the race is to the quick 
learner, Geschwind’s asymmetry could have conferred a selectional advantage 
on Homo loquens as a species, but the brain is plastic and language is learned, 
so mother and family only need to protect any individual symmetric boy or left­
hander for a few extra months until his plastic brain can adapt. 
Environmentally induced critical periods 
While lateralization of the planum temporale might appear genetic to us, to 
your average neuron a flood of testosterone must occur as an environmental 
cataclysm. In fact, numerous exogenous, environmental effects have recently 
been found to affect the developing brain, with indirect consequences for lan­
guage. In 1981, West et al. demonstrated that exposure of rat embryos to ethyl 
alcohol could disorganize development of the hippocampus. In figure 12.2B, 
hippocampal pyramidal cells that normally migrate outward to form a white 
band in the stained preparation of figure 12.2A have clearly failed to migrate 
properly under exposure to ethyl alcohol. It has always been known that chil­
dren of alcoholic mothers were at risk for birth defects, and these were often 
Figure 12.2. 
Ethyl alcohol disorders neuron migration in the hippocampus of fetal 
rats. Normal pyramidal cells form a coherent band in A. In B, after exposure to 
ethanol, the pyramidal cells become physiologically disordered during neuro­
genesis. (West et al. 1981. Reprinted by permission of the American Association for 
the Advancement of Science.) 

176  • 
HOW  THE  BRAIN  EVOLVED  LANGUAGE 
further associated with learning and language disorders, but seeing is believ­
ing. In West et al.’s work, it became clear that the developing brain could be 
physiologically disordered. The teratogenic effects of alcohol upon the devel­
oping brain have been consistently demonstrated when the alcohol is admin­
istered moderately late in gestation. (Earlier poisoning with alcohol can simply 
kill the fetus.) 
Behavioral manifestations of ethanol-induced disorders are often less ob­
vious. In humans, because any linguistic symptoms of fetal alcohol syndrome 
(FAS) cannot be observed until years after birth, such neurological damage 
might be far more prevalent than the number of cases diagnosed and reported 
at birth (currently about 1% of all live births). 
Teratogens, disease, and genetic defects 
Testosterone and alcohol are not the only “environmental” teratogens
2
 to threaten 
the developing fetus. Lead, mercury, common prescription drugs like retinoic 
acid (Retin-A
 or Accutane), and common diseases can all cause birth defects 
and/or learning disorders. Maternal rubella (German measles) causes very seri­
ous birth defects, including deafness. Even maternal varicella (chicken pox) has 
been linked to Reye’s syndrome. Mednick (1994) reported a study of Helsinki 
children in which maternal influenza during the second trimester resulted in 
an exceptionally high incidence of subsequent adolescent schizophrenia, and it 
would not be surprising to find many more-subtle but serious prenatal environ­
mental hazards to language and cognitive development.
3 
Other disorders, for which no environmental cause is apparent, are thought 
or known to be genetic. Fifty years ago, virtually no learning disorders were 
known except Down syndrome and a broadly defined “mental retardation.” 
Today it appears there may be hundreds. The following sections will briefly 
describe a few of these which especially involve language learning. 
Williams Syndrome 
Schizophrenia is a poorly understood and ill-defined disease that often does 
not manifest itself until adolescence, so it is unclear what specific or indirect 
impact influenza might have upon neural development or language develop­
ment, but Williams syndrome may offer an approximate example. Briefly, chil­
dren with Williams syndrome present IQs in the retarded range but normal 
grammar and language. One eighteen-year-old spoke of her aims in life as fol­
lows: “You are looking at a professional book writer. My books will be filled with 
drama, action, and excitement. And everyone will want to read them. . . . I am
going to write books, page after page, stack after stack. I’m going to start on 
Monday” (Reilly et al. 1991, 375). Unfortunately, her plan was delusional. At a 
local, grammatical level of planning and organization, all appears well, but 
globally, her plans and organization are incoherent in a manner vaguely remi­
niscent of schizophrenia or even Wernicke’s aphasia. Wang and Bellugi (1993) 
report that, in contrast to Down syndrome, Williams syndrome is characterized 

WHAT  IF  LANGUAGE  IS  LEARNED  BY  BRAIN  CELLS
?  •  177 
by preservation of prefrontal cortex (where adaptive grammar locates gram­
maticomotor planning) but diminished lateral cortex and interhemispheric 
callosal fibers. This appears to also be associated with the loss of one copy of a 
gene that makes elastin, a protein that is the chief constituent of the body’s 
elastic fibers (Sadler et al. 1994). It would seem that if interhemispheric con­
nections are retarded, development of the arcuate fasciculus might also be 
retarded in Williams syndrome, perhaps sometimes causing symptoms of con­
duction aphasia to also be present. 
Tourette Syndrome 
Whereas Williams syndrome is characterized by loquacious nonsense, Tourette 
syndrome is marked by motor tics and/or terse, involuntary, and often vulgar 
interjections. Recent research ties Tourette syndrome to a hyperactive dopam­
inergic system involving the caudate nucleus (Wolf et al. 1996). Insofar as 
Tourette syndrome might be construed as a kind of hypertonic disorder related 
to obsessive-compulsive disorders (it is exacerbated by dopamimetic drugs like 
L-dopa, which are used to treat Parkinson’s disease), it calls to mind adaptive 
grammar’s earlier suggestion that a tonic articulatory posture might account 
for Klein et al.’s (1994) finding of selective L2 activation in the putamen. 
Tourette syndrome is genetic, but it is incompletely penetrant: identical twins 
develop the disorder to differing degrees, suggesting that nongenetic factors, 
including prenatal influences, can mitigate or exacerbate Tourette syndrome 
(Leckman et al. 1997). 
Offbeat Dysphasia 
Most language-learning disorders report different and subtler symptoms than 
Williams or Tourette syndrome. Gopnik (1990; Gopnik and Crago 1991; Ullman 
and Gopnik 1994) reported on a family in which sixteen of twenty-two mem­
bers in three generations had been independently diagnosed as dysphasic. (The 
term “dysphasia” is typically used to describe a milder aphasia, not of traumatic 
origin.) Unlike in Williams syndrome, the dysphasics’ global, semantic process­
ing appeared to be unaffected. Rather, they produced local errors such as 12.1: 
*The little girl is play with her doll. 
(12.1) 
With a series of tests like the wug test (chapter 9), Gopnik and Crago observed 
that these dysphasics “could not process grammatical features.” In the popu­
lar press, it was widely reported that Gopnik had discovered “the grammar 
gene.”
4 
Cross-linguistic work on agrammatism shows a rather similar pattern (Menn 
and Obler 1990). In wug tests, Broca’s aphasics omit free morphemes like and 
or of and bound syllabic forms like -ing much more frequently than bound, 
nonsyllabic morphemes like -s and -ed. In adaptive grammar, these are the forms 
which are generated “on the offbeat” out of the relation gradient. This analy­

178  • 
HOW  THE  BRAIN  EVOLVED  LANGUAGE 
sis is reminiscent of Leonard’s theory that omitted morphemes will be short in 
duration relative to surrounding units (Leonard 1996),
5
 but it further suggests 
that rhythm disorders may underlie many such linguistic disorders. Rice and 
Wexler (1996, 1253; see also Lahey et al. 1992, 390) criticize Leonard’s account 
on several points: (1) children are more likely to omit is than -ing in sentences 
like He is going to the store, (2) children are more likely to produce the contracted 
is in He’s going than the uncontractible What is it? (3) dysphasic children are 
apt to omit does in sentences like Does he want a cookie? What appears to unify 
Leonard’s and Rice and Wexler’s accounts in these cases is that dysphasic chil­
dren are more likely to omit unstressed (offbeat) grammatical morphemes.
6 
Dyslexia 
When we read, our eyes do not scan the line of text in a smooth, uniform sweep. 
Rather, they jump from fixation point to fixation point in a series of saccades, 
each several words long. Some children have difficulty learning to read and 
are diagnosed as dyslexic. Among dyslexics, there is often a tendency for the 
eye saccades to overshoot their target (Ciuffreda et al. 1985; Eden et al. 1994). 
This is reminiscent of the cerebellar overshoots measured in the Kent et al. 
study (1979) we reviewed at the end of chapter 8, and indeed, the same litera­
ture has often linked dyslexia and cerebellar disorders.
7 
In a closely controlled, double-blind study, Kripkee et al. (1982) reported 
the successful treatment of a small cohort of dyslexics with monosodium gluta­
mate (MSG). As we have seen, glutamate is a ubiquitous excitatory neurotrans­
mitter, so superficially these findings suggest that at least one form of dyslexia 
might result from an underexcited cerebellum. The explanation is complicated, 
however, by the fact that orally ingested glutamate does not readily cross the 
blood-brain barrier.
8
 However, several “glutamate holes” do exist in the blood-
brain barrier, and one leads to stimulation of the pituitary gland of the hypo­
thalamus. The hypothalamus stimulates the sympathetic nervous system, an 
adrenergic system that raises heart and respiration rates. The pituitary gland 
produces about eight major hormones, including ACTH, adrenocorticotrophic 
hormone. ACTH, in turn, stimulates the cortex of the adrenal glands, produc­
ing adrenaline on the other side of the brain barrier, but the nervous system 
carries this information back to the brain, where noradrenaline levels rise in 
response. As we saw in chapter 3 and in Kasamatsu et al.’s results (1979), nora­
drenaline appears capable of increasing a cell’s signal-to-noise ratio and even 
reversing plasticity. 
Kripkee’s clinical observation of one subject followed over many years 
(personal communication) offers a series of deeper clues. The link to MSG was 
first noted after this subject had dinner at a Japanese restaurant, but the sub­
ject had also noted remission of dyslexic symptoms in times of stress: on one 
occasion she reported reading The Amityville Horror (Anson 1977) cover to cover 
with no dyslexic difficulties or reading fatigue. Apparently, adrenaline-raising, 
scary books might be as effective a dyslexia therapy as MSG! 

WHAT  IF  LANGUAGE  IS  LEARNED  BY  BRAIN  CELLS
?  •  179 
But even more appears to be involved than glutamate and noradrenaline. 
Ultimately, Kripkee found that MSG ceased to benefit his longitudinal subject 
at the onset of menopause. Dyslexic symptoms reappeared and were corrected 
only when the subject began hormone replacement therapy. 
Temporal Processing Deficits 
Tallal (1973, 1975) has long studied populations of “language-learning im­
paired” (LLI) children whose grammatical comprehension and general lan­
guage processing are deficient and whose speech discrimination is notably 
deficient in the perception of voice onset time and formant transitions. Tallal 
et al. (1996) reported that a cohort from this population showed general and 
long-term linguistic improvement after four weeks of training on speech whose 
tempi had been digitally slowed and whose formant transitions had been am­
plified by 20 dB. One of the effective training techniques was a competitive 
computer game which rewarded the children with points for correctly process­
ing faster speech segments (Merzenich et al. 1996). 
Attentional Deficit Disorders 
Perhaps the most commonly diagnosed learning disorder is “hyperactivity,” or 
attentional deficit disorder (ADD). Paradoxically, hyperactive children are 
effectively treated with Ritalin
 and other amphetamine-like drugs. These 
drugs raise noradrenaline levels, and in normal people they do not suppress 
but rather cause hyperactivity. Grossberg (1975, 1980) explained this in terms 
of a quenching threshold (QT). An on-center off-surround field has a QT such 
that neural populations activated below the QT are suppressed while popula­
tions activated above the QT are amplified. This is the same process which 
results in contrast enhancement and noise suppression, as discussed in chap­
ter 5. Our ability to “concentrate” or “focus attention” on one particular ob­
ject or event is related to our ability to raise the QT so that all but that object 
of attention is inhibited. In Grossberg’s analysis, “hyperactive” children are 
actually underaroused in the sense that their QT is set too high: nothing ex­
ceeds the QT for attentional amplification, so they are easily distracted by 
objects and events in the environment which they should be able to keep in­
hibited. Eventually, Grossberg’s analysis was accepted, and we no longer say 
that such children are “hyperactive” but rather that they have an “attentional 
deficit.” 
The preceding is but a small sample of the increasing array of learning dis­
orders which twentieth-century science has brought to light. Adaptive gram­
mar doesn’t pretend to have all the answers to these disorders, but it does 
suggest some common themes. For example, as noted in chapter 7, the audi­
tory tract arising from the octopus cells of the cochlear nucleus seems designed 
to detect fast events like the formant transitions of initial consonants. If these 

180  • 
HOW  THE  BRAIN  EVOLVED  LANGUAGE 
cells do not sum to threshold fast enough, adaptive grammar would predict 
deficits similar to those of Tallal’s LLI subjects. But why and how could com­
petitive computer games improve and have a lasting effect upon temporal lin­
guistic performance? 
According to adaptive grammar, Ritalin
, scary books, MSG, and exciting 
computer games could all mitigate temporal processing deficits by stimulating 
noradrenergic and/or glutaminergic arousal systems: a faster system should, 
in general, be able to process faster stimuli. Kasamatsu’s findings suggest that 
increases in noradrenaline could also promote learning and plasticity by oper­
ating directly on synaptic membranes, apparently by facilitating NMDA re­
sponse to glutamate. There may be good reason for professors and brainwashers 
to favor pop quizzes and high-stress environments!
9 
The conditions I have been characterizing as dysphasia are often referred 
to as specific language impairments (SLI). This means that the condition is thought 
to be specific to language, and the existence of such “specific language impair­
ments” is often used to validate the existence of an autonomous “language 
module.” The term “specific language impairment” may be preferable to “dys­
phasia” insofar as it forestalls the prejudicial implication that people who are 
so impaired are significantly impaired in other respects. The preceding example 
of dyslexia attributable to indelicate eye saccades should serve to illustrate that 
such an implication is far from justified. However, it also illustrates the danger 
of assuming that because a disorder specifically affects language, there must 
exist a language-specific module of mind. The cerebellum exerts fine motor 
control over many aspects of behavior; but it is by no means “language-specific.” 
It just so happens that in modern life there is no other behavior so fine, so 
common, and so important as eye saccades during reading. Thus, a cerebellar 
dyslexia may appear to be language-specific, but the cerebellum itself is hardly 
a language-specific “module.” Language is so complex, and it depends so heavily 
upon the adaptive elaboration and interplay of so many different brain sub­
systems, that talk of a monolithic, autonomous language module no longer 
serves as a particularly useful metaphor. 
Morphology 
From chapter 8 onward, rhythmicity and morphology have emerged as cen­
tral players in the evolving organization of language. Morphology has been 
closely examined in the first-language-learning literature since Brown 1973, 
but it has recently come under increasing attention. In part, this interest has 
been driven by connectionist models of learning English regular and irregu­
lar past tense forms (Rumelhart and McClelland 1986a; MacWhinney and 
Leinbach 1991; Plunkett 1995; see Pinker and Prince 1988, 1994, for critiques), 
and in part it has been driven by studies of dysphasia and specific language 
impairment. 
Whereas Tallal’s LLI subjects were impaired at time scales of 0–100 ms, the 
morphological disorders of Gopnik’s dysphasics involved grammatical mor­
phemes like -ing, which may have durations of 200 ms and more. In chapter 9, 

WHAT  IF  LANGUAGE  IS  LEARNED  BY  BRAIN  CELLS
?  •  181 
I posited two rhythm generators, one for syllables, on the former time scale, 
and one for metrical feet, on the latter time scale. These rhythm generators 
might be differentially involved in the two disorders. However, Gopnik’s sub­
jects also had significant phonological impairments, and Tallal’s subjects may 
also have had grammatical impairments like those of Gopnik’s subjects, so it is 
possible that small temporal deficits in motor and sensory processing like 
Kripkee’s and Tallal’s can become compounded to also produce dysphasic 
deficits. Even dyslexia is frequently linked to subtle morphological and pho­
nological deficits. Eden et al. (1994) noted an inverse correlation between 
dyslexia and the ability to process pig latin—which the reader will now recog­
nize as a metathetic skill. 
A second theme underlying the preceding results is that of rhythm and 
cerebellar function. Until recently, the cerebellum has been widely discounted 
as a player on the linguistic and cognitive stage. The cerebellum’s effects are 
often subtle (Holmes [1939] reported that people can compensate adequately 
for even severe cerebellar damage), and its obvious role in motor behavior fixed 
its theoretical position on the machine side of the Cartesian mind/machine 
dichotomy. My theory suggests that motor rhythm is, like gravity, a weak but 
pervasive force, and recent research has begun to also associate dysmetria with 
conditions like dyslexia (Nicolson et al. 1995; Ivry and Keele 1989). 
Autosomal dominance 
Rather remarkably, the distribution of symptoms found throughout three gen­
erations of Gopnik’s subject family quite conclusively implied that the dis­
order was autosomally dominant (Hurst et al. 1990). Similarly, Kripkee and others 
have reported evidence of autosomal dominance for dyslexia. There is also evi­
dence to suggest autosomal dominance in Williams syndrome, and Tourette 
syndrome is known to be autosomally dominant. Autosomal dominance means 
that a genetic trait is neither sex-linked nor recessive. Other things being equal, 
it means that these language disorders should become increasingly prevalent 
in the human gene pool! 
Put differently, it seems likely that we are all language-impaired. To be sure, 
some of us are more impaired than others. As we noted in the case of Tourette 
syndrome, disorders can exhibit incomplete penetrance: pre- and postnatal 
factors can partially block expression of even a dominant trait. In most cases, 
I believe our brains simply learn to compensate for our disabilities. We may need 
to learn as much to survive our disabilities as we need to learn to survive our 
environment. 
Postnatal Language Learning 
Despite the gauntlet of teratogens, diseases, and genetic anomalies we encoun­
ter, we mostly survive, and by birth much remarkable neural development has 
occurred. After the augmentation of the planum temporale, the arcuate fas­

182  • 
HOW  THE  BRAIN  EVOLVED  LANGUAGE 
ciculus and auditory pathways from ear to cortex have been laid down and have 
begun to myelinate.
10
 By about five months, the child’s full complement of 
neurons has been produced by mitotic division. No more will be created. For 
the next few years, many of these cells will continue to grow and form connec­
tions among themselves, but at the same time many of them, perhaps even most 
of them, will begin to die off in a massively parallel competition to the death. 
This phenomenon is sometimes discussed along with “programmed cell 
death,” or apoptosis, the theory that every cell of the human body is programmed 
to die. Apoptosis is an especially important theory in areas like oncology, where 
it explains tumors as a programmatic failure that makes cancerous cells “im­
mortal.” In the case of neurons, however, it seems that this early cell death is 
less programming than the result of neurons failing to make the right connec­
11
tions; I will call it neuroptosis.
Much of what we know of human neural connectivity was learned when 
the injury and death of a first population of neurons (as in amputation) caused 
atrophy and death in a second population which it innervated. Such atrophy-
and-death by inactivation was the technique used in Hubel and Wiesel’s early 
studies, and it would also explain a case of human neonatal neuroptosis first 
observed by Ramón y Cajal. He discovered an entire class of neurons in the 
most superficial layer of fetal human cortex (“horizontal cells”) that are not 
present in adult cortex. These cells’ dendrites are oriented parallel to the cor­
tical surface (i.e., horizontally), giving the appearance that their outward de­
velopmental migration ends in a “crash” against the skull. Crashed against the 
skull and misoriented for resonance in cerebral cortex, where neurons are 
optimally arranged in vertical barrels, they atrophy and die. 
Neuroptosis was first well documented in modern science in chick embryos, 
where up to 50% of the spinal cord neurons were found to die off between six 
and nine days after fertilization (Hamburger 1975). Similar neuroptosis was 
then rapidly found in other species, with Lund et al. (1977) finding a regres­
sion in the spininess of primate (macaque) pyramidal neurons between age 
eight weeks and two years. It is now clear that a large-scale sacrifice of neurons 
occurs as well in the human fetus and then continues at a decreasing rate for 
some months and years after birth. It is as if the fortunate neurons get into 
college, get connected, and thrive; the unconnected lead marginal lives and 
die young. 
The child’s first year is not totally prelinguistic. Considerable receptive 
language learning clearly occurs, but the infant’s capacity for motoric response 
is limited, and only the most careful and ingenious experiments can assay the 
extent of this receptive learning. In one such experiment, Jusczyk and Hohne 
(1997) read 15 eight-month-old infants three stories containing novel words 
ten times over a two-week period. At the end of the period, the infants showed 
a small but statistically significant tendency to turn their heads in attention to 
words which they had heard in the stories. 
Mothers know their children are listening, and their speech is also care­
fully tuned to the developing infant’s language-learning needs. In chapter 6 
we noted that the mother’s and child’s voices have poorly defined vowel for­

WHAT  IF  LANGUAGE  IS  LEARNED  BY  BRAIN  CELLS
?  •  183 
mants because they have fewer harmonics than adult male voices, and we won­
dered how this could be adaptively conducive to language learning. It turns 
out, however, that “motherese” and “baby talk” are especially characterized by 
precise vowels (Bernstein-Ratner 1987) and wide intonation contours (Fernald 
and Kuhl 1987). Thus, as mother’s fundamental frequency swoops up and down 
in careful but expressive intonation, the harmonics also “swoop” through the 
vocal tract filter, filling formant resonances at all frequencies, not just at the 
discrete harmonics of a monotonous voice. 
In the first year of life, Piaget found motor intelligence developing from 
circular reactions: the child can see her hand moving, and from this observation, 
she learns that she can move her hand.
12
 From this, in turn, develops a sense 
of self and an evolving intelligence, leading through “egocentricity” to the social 
awareness of the adult. 
Grossberg (1986; Grossberg and Stone 1986) applied this Piagetian “schema” 
to babbling and the ontogenesis of speech (figure 12.3). In this model, infant 
babbling is just a random motor movement until some one such motor gesture 
(e.g., the babble mamamamama) becomes regularly paired with a sensory pat­
tern (e.g., the visual presence of Mama). In figure 12.3, this occurs between F
S 
and F
M
. Adaptive grammar associates this pathway with the arcuate fasciculus. 
The motor pattern and the sensory pattern resonate across the arcuate fascicu­
lus in short-term memory until, by equations 5.3 and 5.4, a long-term memory 
of the resonance forms.