Oxford university press how the br ain evolved language

Figure 5.4. 
(a–b) A cerebral gyrus simulated by equations 5.2–5.4. (c–e) A cerebral 
gyrus simulated by equations 5.2–5.4. (f) A gyrus of macaque striate cortex radio­
graphically labeled by tritium. (LeVay et al. 1985. Reprinted by permission of the 
Society for Neuroscience.) 
Edge detection 
In figure 5.4b, inputs continue to be applied through t = 1, and a resonant 
pattern begins to develop. Edge detection emerges as another inherent property 
of the ART system: the non-edge nodes at [10 11] and [10 13] are actively in­
hibited on both flanks (by each other, as well as by the edge nodes [10 9] and 
[10 15]). On the other hand, the edge nodes are each actively inhibited only 

84  • 
HOW  THE  BRAIN  EVOLVED  LANGUAGE 
by nodes on one interior flank (by [10 11] or [10 13]). Consequently, the edges 
are less inhibited, more active, and more perceptually detectable. 
Normalization 
A fourth property of on-center off-surround anatomies is normalization. Con­
sider as an example the intensity of speech sounds. For human hearing, the 
threshold of pain is somewhere around 130 dB. What happens at a rock con­
cert, where every sound may be between 120 dB and 130 dB? In the on-center 
off-surround anatomy of auditory cortex, if one is not first rendered deaf, a 
frequency component at 130 dB will inhibit its neighbors. A neighboring fre­
quency component at 120 dB can thus be inhibited and perceived as if it were 
only 80 or 90 dB. In this way, an on-center off-surround anatomy accomplishes 
a kind of automatic gain control—normalization that prevents the system from 
saturating at high input levels. 
On-center off-surround anatomies accomplish a similar kind of normal­
ization in vision. At sunset, daylight is redder than it is at noon. Nevertheless, 
at sunset we still perceive white as white, not as pink. In this case, an on-center 
off-surround perceptual anatomy keeps the world looking normal even as its 
optical characteristics physically change. 
Rebounds 
Figure 5.4c displays NSA applied at t = 2 to all nodes in the gyrus. This is analo­
gous to the “flashbulb effect” discussed with reference to figure 5.3. After NSA, 
at t = 3 (figure 5.4d), a rebound occurs: the target cells, which previously were 
“on” relative to their surround, are now “off.” Note that the rebound does not 
only occur on the local scale of the original target barrels. A larger scale, self-
similar rebound can also occur over the left and right hemifields of the gyrus 
(fields [x 1–20] vs. [x 21–48]). Rebounds occur on large, multicellular scales 
just as they occur on smaller, cellular scales. This capacity for fieldwide rebounds 
is critical to preserving long-term memory invariance. 
Complementation and long-term memory invariance 
Long-term memory is, by definition, invariant: A pattern once learned (like 
that in figure 5.4a) should not be easily forgotten. But amid all the busy, buzz­
ing, resonant neural activity suggested by figure 5.4, what is to prevent a re­
membered pattern like that of figure 5.4a from being overwritten by other, 
conflicting inputs? Rebounds are the answer to this cognitive problem, too. 
Suppose that I ask you to learn the pattern in figure 5.4a, but after you have 
studied it for a little while, I change my mind and say, “No, stop. Now learn 
this second pattern.” In later chapters we will see how my No instruction causes 
NSA and a general rebound, putting your mind in the tabula-not-quite-rasa state 
of figure 5.4d. If, while you are in this new state, I present you with a new pat­

ADAPTIVE  RESONANCE 
•
85
 
tern to learn, the original target nodes will be inactive. By equation 5.2, synap­
tic learning at those nodes will also be inactivated. (Let the rebounded, inhib­
ited target nodes, x
r
, in figure 5.4e have short-term memory (STM) activation 
levels of 0, i.e., x
r 
= 0. Then by equation 5.2, Ex
i
x
r
 = 0, so all z
ir
 (0.) In this way, 
rebounds complement memory, partitioning it and preventing new information 
from overwriting old, long-term memories. 
Ocular dominance columns 
If the resonance begun in figure 5.4a–b is allowed to resonate without conflict­
ing inputs, the gyrus eventually achieves the pattern of figure 5.4e. Figure 5.4e 
bears a striking resemblance to ocular dominance columns in primary visual (stri­
ate) cortex (figure 5.4f). 
Wiesel and Hubel (1965; Wiesel et al. 19974) sutured one eye of newborn 
kittens closed. After several months, this caused neocortical cells that were wired 
to the sutured eye to become permanently unresponsive. Similar suturing of 
adult cats had no such effect, so Wiesel and Hubel proposed that there existed 
a critical period during which vision had to become established by experience. 
Subsequent radiographic staining techniques allowed Wiesel, Hubel, and Lam 
(1974) and others to make dramatic pictures showing that eye-specific striate 
cortex cells were arranged in stripes, or “columns.” Figure 5.4f is one such pic­
ture, in which white stripes are neurons responding to one eye (not sutured 
but radiographically labeled by tritium; LeVay et al. 1985). Like other sensory 
systems, the visual system had been known to exhibit a retinotopic mapping 
from peripheral receptors in the eye up through cortex. Ocular dominance 
columns appeared as one more remarkable instance of such topographical 
mapping. (The tonotopic mapping of the auditory system will figure prominently 
in our next several chapters on speech and speech perception.) 
Hubel and Wiesel’s studies exerted a broad influence in science, and Lenne­
berg (1967) proposed that a similar “critical-period” explanation could be ex­
tended to language. Coupled with Chomsky’s speculations on the innateness of 
language, Lenneberg’s thesis suggested that there existed a detailed genetic plan 
for grammar, just as there could be supposed to exist a detailed genetic plan in 
which segregation of thalamocortical axons formed the anatomic basis for detailed 
ocular dominance columns. Because only about 10
5
 human genes are available 
to code the 10
8
-odd axons innervating striate cortex (not to mention Broca’s area, 
Wernicke’s area, and all the rest of the human brain), this suggestion was never 
adequate, but in the absence of better explanations it was widely accepted. 
We will return to these issues of critical periods and neuronal development 
in later chapters. For now, it remains only to establish that the similarities be­
tween our model gyrus and real neocortex are not fortuitous. To this end, note 
that the similarities between figure 5.4e and figure 5.4f are not only impres­
sionistic and qualitative but also quantitative: the diameter of the stripes in 
figure 5.4e is approximately two barrels—the radial extent of a barrel’s inhibi­
tory surround in the model gyrus. It also happens that the width of Wiesel and 

86  • 
HOW  THE  BRAIN  EVOLVED  LANGUAGE 
Hubel’s ocular dominance columns is 0.4 mm—approximately the radial ex­
tent of cerebral inhibitory cells’ inhibitory surround.
6 
XOR 
The potential advantages of massively parallel computers have been vaguely 
apparent for quite some time. In fact, the first large-scale parallel computer, 
the 64,000-element Illiac IV, became operational in the mid-1960s. But before 
the invention of microchips, parallel machines were prohibitively expensive, 
and once built, the Illiac IV proved extremely difficult to program. The lead­
ing parallel-computing idea of the day was Rosenblatt’s perceptron model (Rosen­
blatt 1958, 1959, 1961), but in 1969 Minsky and Papert’s widely influential book 
Perceptrons (1969; see also Minsky and Papert 1967) claimed to prove that percep­
trons were incapable of computing an XOR (see below). As a result, they ar­
gued, perceptrons were incapable of calculating parity, and therefore incapable 
of performing useful computational tasks.
7
 Because of this argument, XOR has 
figured prominently in subsequent parallel-computing theory. We will return 
to this issue, but for the present, consider only that dipoles calculate XOR. 
XOR, or “exclusive OR,” means “A or B but not both.” Formally, it is a 
Boolean logical operation defined for values of 1 and 0 (or true and false). 
For the four possible pairs of 1 and 0, XOR results in the values given in table 
5.2. XOR is 1 (true) if A or B—but not both—is 1. This is the same function
that gated dipoles computed in figure 5.3 and 5.4. Grossberg (1972a, passim) 
found that gated dipoles compute XOR ubiquitously in the brain, as an essen­
tial and natural function of agonist-antagonist relations and lateral inhibition. 
Calculating parity is no longer thought essential to computation, but as we have 
seen and as we shall see, dipoles and XOR are essential to noise suppression, 
contrast enhancement, edge detection, normalization, long-term memory in­
variance, and a host of other indispensable properties of cognition. 
Opportunistic Learning with Rebounds 
The rate of neurotransmitter release from the presynaptic terminal is not con­
stant. When a volley of signal spikes releases neurotransmitter, it is initially 
released at a higher rate. With repeated firing, the rate of release of neurotrans­
mitter decreases. As suggested previously, this constitutes synaptic habituation 
(see also Klein and Kandel 1978, 1980). 
TABLE 5.2. 
Truth table for exclusive OR (XOR). 
A: 
1 
1 
0 
0 
B: 
1 
0 
1 
0 
XOR 
0 
1 
1 
0 

ADAPTIVE  RESONANCE 
•
87
 
A plausible mechanical explanation for this habituation is that while a knob 
is inactive, neurotransmitter accumulates along the synaptic membrane (see 
figure 3.7). When the first bursts of a signal volley depolarize this membrane, 
the accumulated transmitter is released. The synapse receives a large burst of 
neurotransmitter and a momentary advantage in dipole competition. There­
after, transmitter continues to be released, but at a steadily decreasing rate. 
Meanwhile, the inhibited pole of a dipole is dominated, but it is not slavish. 
All the while it is dominated, it accumulates neurotransmitter reserves along 
its synaptic membranes, preparing itself for an opportunistic rebound. 
Rebounds make for opportunistic learning in a manner reminiscent of 
Piagetian accommodation (Piaget 1975). Learning implies the learning of new 
information, so if old information is not to be overwritten when the new infor­
mation is encoded, a rebound complements and repartitions memory to ac­
commodate the new information. 
Expectancies and Limbic Rebounds 
A few pages back it was suggested that a rebound could be caused by a teacher 
saying, “No, wait. Learn a different pattern.” But what if there is no teacher? 
For billions of years, intelligence had to survive and evolve in the “school of 
hard knocks”—without a teacher. How could rebounds occur if there were no 
teacher to cause them? 
In figure 5.4b, we saw secondary, resonant fields form, encoding the input 
pattern presented in figure 5.4a. In a richer environment, these secondary 
resonant fields encode not only the immediate input but also, simultaneously, 
the associated context in which input occurs. In the future, the learned z
ij
 traces 
of context alone can be sufficient to evoke the target pattern. These second­
ary fields are contextual expectancies, or, as Peirce put it in 1877, “nervous asso-
ciations—for example that habit of the nerves in consequence of which the 
smell of a peach will make the mouth water” (9). 
In 1904, Pavlov won the Nobel Prize for discovering (among other things) 
that associations could be conditioned to arbitrary contexts. For example, if 
one rings a bell while a dog smells a peach, one can later cause the dog’s mouth 
to water by just ringing the bell. In 1932, Tolman gave such abstract contexts 
the name we use here—expectancies. Grossberg’s suggestion (1980, 1982a) was 
that failed expectancies could trigger rebounds, thereby making it possible for 
animals and humanoids to learn new things during the billion years of evolu­
tion that preceded the appearance of the first teacher. 
Evoked Potentials 
As we have seen, the action potentials of neurons are electrical events, and in 
1944, Joseph Erlanger and Herbert Gasser received the Nobel Prize for devel­
oping electronic methods of recording nerve activity. One outgrowth of their 

88  • 
HOW  THE  BRAIN  EVOLVED  LANGUAGE 
work was the ability to measure evoked potentials—the voltage changes that a stimu­
lus evokes from a nerve or group of nerves. Allowing for variation among differ­
ent nerve groups, a fairly consistent pattern of voltage changes has been observed 
in response to stimuli. During presentation of a stimulus sentence like 
Mary had a little lamb, its fleece was white as coal 
(5.5) 
CNV 
P100 P300 N400 
a series of voltage troughs and peaks labeled CNV, P100, P300, and N400 may 
be recorded by scalp electrodes positioned over language cortex. CNV, or con-
tingent negative variation is a negative voltage trough that can be measured as a 
usual correlate of a subject’s expectation of a stimulus, for example, sentence 
5.5. A positive peak, P100, can then be recorded 100 ms after stimulus onset,
when a stimulus is first perceived. If, however, the stimulus is unrecognized— 
that is, if it is unexpected—a P300 peak is reliably evoked, approximately 300 ms 
after stimulus onset. The rapid succession of syllables in a sentence like 5.5 can 
obscure these potentials on all but the last word, but in cases of anomalous 
sentences like 5.5 (Starbuck 1993), there is a widespread N400 “anomaly com­
ponent” which is clearly detectable over associative cortex. 
Grossberg’s theory (1972b, 1980; Grossberg and Merrill 1992) is that P300 
corresponds to a burst of NSA that is triggered by the collapse of an expecta­
tion resonance. More specifically, Grossberg (1982a) suggests that old knowl­
edge and stable expectancies have deep, subcortical resonances extending even 
to the hippocampus and limbic system. If events in the world are going accord­
ing to the mind’s plan, then heart rate, breathing, digestion, and all the other 
emotions and drives of the subcortical nervous system are in resonance with 
the perceived world, but a disconfirming event, whether it be the unexpected 
attack of a predator, a teacher’s No! or the simple failure of a significant ex­
pectancy causes this harmonious resonance to collapse. This collapse unleashes 
(disinhibits) a wave of NSA, causing a rebound and making it possible for the 
cerebrum to accommodate new information. 
Grossberg’s P300 theory is also compatible with Gray’s “comparator” theory 
of hippocampal function, and together they give a satisfying explanation of 
certain facts discussed in chapter 4, like HM’s anterograde amnesia. Accord­
ing to this Gray-Grossberg theory, HM could not learn new things because, with 
a resected hippocampus, he could not compare cerebral experience against 
his subcerebral emotional state and so could not generate rebounds. Without 
rebounds, his cerebral cortex could not be partitioned into active and inactive 
sites, so inactive sites could not be found to store new memories. 
Although HM could not remember new names and faces, he could retain 
certain new, less primal long-term memories such as the solution to the Tower 
of Hanoi puzzle. By the Gray-Grossberg theory, HM might have been able to 
retain these memories because nonemotional experiences are relatively non-
limbic and may therefore depend to a lesser extent on limbically generated 
NSA. As we saw from figure 5.3, NSA is not the only way to trigger a rebound. 

ADAPTIVE  RESONANCE 
•
89
 
Turning down inputs by shifting attention (or closing the eyes) can generate a 
rebound just like turning up inputs. This is why we are often able to solve a 
problem by “sleeping on it.” Turning off the inputs that have been driving our 
cognitive processes allows alternative hypotheses to rebound into activity and 
successfully compete for LTM storage. 
Grossberg’s P300 hypothesis suggests that Gray’s comparator theory might 
be tested by evoked potentials, but unfortunately, the hippocampus and other 
key limbic centers lay deep within the brain, inaccessible to measurement by 
scalp electrodes and even by sophisticated devices like magnetoencephalograms 
(MEG scans). On the other hand, measurement techniques which can “go 
deep,” like positron emission tomography (PET) scans and functional magnetic reso-
nance imaging (fMRI) scans, resolve events on a scale of seconds at best—far 
off the centisecond timescale on which evoked potentials occur. 
Sequential parallel search 
Repeated rebounds can enable a sequential parallel search of memory. If a 
feedforward input pattern across some field F
1 
does not match (that is, reso­
nate with) an expected feedback pattern across F 
2
, a burst of NSA can cause a 
rebound across F 
2
, and a new feedback pattern can begin to resonate across 
F 
2
. If this new F 
2
 expectancy does not match the input, yet another rebound 
can cause yet a third feedback pattern to be instantiated across F 
2
. This pro­
cess can repeat until a match is found. 
Peirce (1877) described this process as abduction, which he contrasted with 
the classical logical processes of induction and deduction. He exemplified it 
with the story of Kepler trying to square Copernicus’s theoretical circular or­
bits with the planets’ occasional retrograde motion. Kepler tried dozens of 
hypotheses until he finally hit upon the hypothesis of elliptical orbits. While 
Kepler’s trial-and-error abductive process was not logical in any classical sense, 
it did reflect the logic of expectancies, trying first the expected classical, Eu­
clidean variants of circular orbits before finding resonance in the theory of 
elliptical orbits. 
Such searches are not random, serial searches through all the 10
7,111,111
-odd 
patterns in the mind. Rebounds selectively occur in circuits which have mini­
mal LTM (minimal neurotransmitter reserves) and maximal STM activity. By 
weighing experience against present contextual relevance, the search space is 
ordered, and—unlike a serial computer search through a static data structure— 
abduction quickly converges to a best resonance. 
In this chapter we have modeled cerebral dynamics using only a handful of 
basic differential equations. These equations nevertheless enabled us to build 
a computer simulation which exhibits subtle cognitive dynamics and surpris­
ing similarities to real mammalian neocortex. Many of these dynamic proper­
ties are especially evident in language phenomena, to which we will turn in 
chapter 7. First, however, chapter 6 must give the reader a quick survey of the 
physics and physiology of speech and hearing. 

90
•
HOW THE BRAIN EVOLVED LANGUAGE
• 
S 
I 
X
• 
Speech and Hearing
In the previous chapters we reviewed the central nervous system and introduced 
adaptive resonance theory (ART) as a system for modeling it. Most of the data 
presented in those chapters dealt with vision. Language, however, deals pri­
marily with sound. Therefore, in this chapter we will first consider some essen­
tial, physical facts of speech. We will then see how these physical data are 
processed by the ear and the auditory nervous system into the signals that the 
cerebrum ultimately processes as language. 
Speech 
Periodic sounds 
Speech may be divided into two types of sounds: periodic and aperiodic. These 
types correspond roughly to the linguistic categories of vowel and consonant. 
We will focus on these categories because our first linguistic neural networks 
(in chapter 7) will model how we perceive them. 
Vowels are the prototypic periodic speech sounds. They originate when 
air is forced from the lungs through the glottis, the opening between the 
vocal cords of the voice box, or larynx (figure 6.1).
1
 When air flows between 
them, the vocal cords vibrate. In figure 6.2, a single string, which could be a 
violin string, a piano string, or a vocal cord, is shown vibrating. 
Each time the string in figure 6.2 stretches to its right, air molecules are 
set in motion to the right. They bump against adjacent molecules and then 
bounce back to the left. In this manner a chain reaction of compressed “high­
pressure areas” is set in motion to the right at the speed of sound. (Note that 
molecule A does not wind up at position Z 1 second later. Rather, one must 
imagine that molecule B bumps molecule C to the right. Molecule B then 
bounces back and bumps molecule A to the left. Eventually, Y bumps Z. When 
90 

SPEECH  AND  HEARING 
•
91
 
Figure 6.1. 
The larynx. 
the string springs back to the left, alternating low-pressure areas are formed 
between the high-pressure areas. The resulting sound waveform spreads like 
the waves from a pebble dropped in a still pond. If we plot these high- and low-
pressure areas (figure 6.2), we get a sinusoidal (sine-wave-shaped) sound wave. 
These waves are commonly seen when a microphone transduces high and low 
air pressure waves into electrical signals displayed on an oscilloscope. 
The ordinate of the plot (AMP in figure 6.2) shows how the pressure rises 
and falls. This measure of rise and fall is the wave’s amplitude. It corresponds 
to the sound’s intensity or loudness. (The human ear does not hear very 
low or very high sounds, so technically, “loudness” is how intense a sound 
seems.) In figure 6.2 the wave is shown to complete two high-low cycles per 
second. The wave’s measured frequency in figure 6.2 is therefore two hertz 
(2 Hz).
2 
Every string has a natural fundamental frequency (f
0
), which depends upon 
its length and elasticity. When a string of given length and elasticity vibrates as 
a whole, as in figure 6.2, it produces the first harmonic (H
1
). But if we divide 
the string in half by anchoring its midpoint, as in figure 6.3, each half-string 
will vibrate twice as fast as the whole. The plot of the resulting high- and low-
pressure areas will be a wave with twice the frequency of f
0
. Figure 6.3 is drawn 
to show such a wave at 4 Hz, twice the fundamental frequency in figure 6.2. In 
musical terms, H
2
 sounds an octave higher than the fundamental frequency, 
f
0
. Note, however, that each half-string in figure 6.3 moves less air than the whole 
string of figure 6.2, so the amplitude of H
2
 is less than that of H
1
. 

92  • 
HOW  THE  BRAIN  EVOLVED  LANGUAGE