[comp.ai] Genesis of language

msellers@mntgfx.mentor.com (Mike Sellers) (12/03/87)

[I've just recently gotten my posting powers back, so this may seem somewhat
late.  However, the discussion doesn't seem to have progressed too far in
terms of answering some of the basic questions involved here, so I thought
I'd go ahead and throw in some neurological data.  I've included salient 
portions of the original article that started this whole thing, along with 
my comments.  
I'd appreciate comments, as I've not seen much in the way of cognition or 
linguistics from a neurological point of view on the net.]

Mike Glantz wrote an article that ended with:
>  Does anyone have any concrete information about human brain physiology
>  which would favor the completely ``physiological'' hypothesis of
>  linguistic capability over the ``sociological/anthropological''
>  explanation, or which would shed any other light on the question? 

  I haven't seen much in the way of concrete information about these 
questions on the net, so I'm posting what I know.  I think many of the people 
interested in problems like this one would do well to become more familiar 
with recent neurological findings; while they often follow what you might 
assume or intuit to be true, the human brain is often stranger and more 
elegant that you would imagine.  
  The rest of the references here are from Mike Glantz's original article.

>  Much discussion about neural networks carries the implication that it
>  is a human brain researchers are hoping, ultimately, to simulate, and
>  that a successful simulation will exhibit human linguistic capability.
>  This is certainly an admirable and worthwhile, if ambitious, goal. But
>  current models don't seem to have any features which would distinguish
>  a human brain from, say, a cat's brain (I realize this is very early
>  days - no criticism intended). This will eventually have to be dealt
>  with. 

  More precisely: current connectionist models are much closer to the brain
of the Aplysia (sea hare, a type of sea slug), or even the planaria's ganglia,
than they are to the human brain, both in terms of absolute neural complexity
and internal symbolic structure.  Both the amount of neural structure (whether 
biological or synthetic in origin) and the synaptic and symbolic organization 
of that structure are important to an understanding of what is happening and 
how it happens.  (I am using the work 'symbolic' here to denote any software-
like components of the neural organization; this level of complexity may 
derive some of its attributes from the underlying physical structure --what 
Pylyshyn calls the 'functional architecture'-- but the specific function of 
the architecture is not derivable by examining the structure itself.)
  Little is known about how and why the human brain organizes (lateralizes)
itself on a neural or nuclear (groups of neurons) level as it does.  This
knowledge is crucial to performing any sort of artificial simulation of human 
linguistic capabilities.  What is known, however, can shed some light on many
of the questions being bandied about in this discussion.

>  One possible explanation for why humans have language and cats don't is
>  that there may be one or more physiological structures unique to the
>  human brain, other than its larger capacity, which make language
>  possible. This is the most obvious explanation that comes to mind, and
>  is perfectly reasonable, although we haven't yet identified which
>  structures these are, or what roles they might play. 

  On a large scale, it is known that two areas of the brain are specific to
linguistic ability.  These are Broca's and Wernicke's areas, which appear 
in the left frontal and temporal cortices of most adult humans.  They do
not appear in other animals.  Some humans develop with these areas in other
places (i.e. about 30% of all left-handers lateralize with these areas in 
the right hemisphere, and a few people seem to have speech control resident 
in both hemispheres), but with only a few pathological exceptions, they do 
appear in all human brains.  It is probable that the amount of cortical mass
does have to with the ability to develop areas like Broca's and Wernicke's:
cortical real estate is expensive, so having areas with functions like 
linguistics probably depends on having enough 'other' mass to devote to 
everything else the organism needs to be doing.

>  But another possibility is that maybe the larger brain capacity is
>  sufficient, but that language is possible only after certain
>  ``internal'' or ``symbolic'' structures are built on top of the
>  physiological base. This building occurs during infancy and early
>  childhood, and the resulting structures can be considered to be part of
>  the human brain, every bit as real as the physiologically observable
>  features. 

  There are structural differences between Broca's and Wernicke's areas and
the rest of the cerebral cortex, and identifiable connections between these
two areas (called the Arcuate fasciculus, I believe) as well.  Broca's area
is adjacent to the motor cortex, and controls facial expression, phonation,
etc., while Wernicke's area controls comprehension, sentence construction,
etc.  Many tests and case studies have shown how integral these two areas 
are to our creation and comprehension of speech.  On the other hand, these
areas do not appear to be differentiated from the rest of the cortex at birth.
Some areas, such as the visual, sensory, and motor cortices, are already 
well developed and dedicated to their specific function at birth, even though
the brain is still rapidly growing at this point (some estimates put the rate 
of growth at 100,000 new neurons per minute!).  Other areas, such as most 
of the pre-frontal and temporal lobes, appear to be 'blank' at birth.  That 
is, the neural and glial structures are present, but no specific function has 
been assigned to or adopted by that area.  
  What is interesting is what happens beginning just before birth, and for 
several years afterwards: neurons die in droves.  It seems that each neuron
sends out many (thousands) of afferent and efferent fibers to other neurons
(how each knows which and how many to send out initially is still a mystery).
These fibers will eventually become dendrites (for input) and axons (for
output).  What then happens is that those fibers that are used (stimulated)
become thicker and stronger, and in the case of dendrites, put out smaller
hair-like fibers.  Those fibers that are not used die off, severing the
previously made connection.  If enough fibers from a cell are not used, the
cell itself dies.  Probably what is happening is that those fibers that are
used get preferential supplies of mitochondria and Golgi complexes, leaving
others to wither from disrepair; if too few fibers are used, the cell itself
does not maintain enough mitochondria and Golgi complexes to keep it going,
and so it dies.  The upshot of this is a sort of 'survival of the fittest'
among neurons: those that are used the most survive, while others die. 
Since connections between neurons in the brain is most of what matters (if
not all that matters), this has a profound effect on the developing organism
as a whole.  This process of neural death peaks out in humans at around 
5 years old, I believe, and generally ends by the time we are 7-10 years old.
It is unclear how much neural death and/or regeneration takes place after 
this point, but it is clearly over on any large scale by this time.  In some
areas of the brain, 10-20% percent of the neurons die (as in the visual
system, where the error rate for initial fibers seems to be as low as 5%,
even among billions of possible connections), while in other areas (such as
the pre-frontal cortex), up to 85% of the initial population die off.  It
is probably safe to make a connection between those areas that have high
incidence of neural death and those that are affected by the environment and
other non-biological factors.  The pre-frontal, temporal, and parietal lobes
(where we do much of our "thinking", associating, comprehending, remembering,
speaking, pattern matching, etc.) are all severely affected by the neural
death.

>  [...]
>  The principal hypothesis, here, is that, given sufficient relative
>  brain capacity, and the appropriate socialization process, any
>  individual of another species (a porpoise, for example) could acquire
>  linguistic ability. 

I don't think so.  What is involved in the development of the brain is more
than just socialization; it also has to do with feedback with sensory and 
motor targets in the body (if a mouse has no whiskers on one side, all 'those'
neurons will connect up with the whiskers on the other side) as well as with
evolutionary trends.  Clearly the environment (and thus socialization) play
a large role in how brain develops, as studies with rich/nominal/deprived
environments have shown, but this is not the only factor.  Even brain size
(or more accurately, central nervous system weight to body weight ratio)
is not necessarily a limiting factor.  Porpoises, for example, though they
have a CNS to body ratio similar to humans, have little 'blank' space in
their brains that could take up tasks like high-order association or 
lingusitics.  That they do not possess the organs for speech further compounds
the problem.  
  While animals in 'enriched' developmental environments will end up with 
thicker, denser cortices, better dendritic connections between neurons,
and seemingly more intelligence, they will not spontaneously start using
portions of their brains for previously unknown tasks (at least, not so far
as we know :-) ).  Thus a mouse will not develop a more complex mouse-language
or the ability to perform previously undo-able tasks after being raised in
an enriched environment.  It may do tasks that other mice can do better than
many of them, but it will not start doing really new things.  (While this
might lead you to believe that *no one* could come up with new cortical 
functions, keep in mind that all studies done so far do not take in to account
evolutionary time periods.  This could make a large difference.)  

>  [Aside: It is known that the human brain (and that of other mammals, as
>  well) undergoes physiological changes during the period of infancy and
>  early childhood. It is possible that the initial acquisition of
>  linguistic skills can only occur effectively during this period, during
>  which time these physiological changes are significantly ``molded'' by
>  the socialization process, where certain ``symbolic'' structures
>  actually become ``wired in''. If this were the case, then the period
>  during which basic linguistic ability can be acquired would be limited
>  to this ``crystallization'' period, which is possibly much longer in
>  humans than in other mammals. We would then have to amend the
>  hypothesis to read: given sufficient brain capacity and a sufficiently
>  long ``crystallization period'' etc. It then remains (among other
>  things) to determine the exact nature of this ``crystallization'', and
>  incorporate a sufficiently long duration of this in a computer model. 

  I have discussed briefly the period of development/molding that comes 
about in the CNS by the process of massive neural genesis, followed by 
massive neural death.  This is almost certainly responsible for much of the
high learning rate seen in human children.  Once the brain is relatively
stabilized (after age 8 or so), it may be that all subsequent learning is
accomplished with intra-neuron changes and changes in synaptic weights.
It is probable that some neural change occurs in response to learning in
adults, though nothing like what is seen in children.  This is also probably
the source of the 'crystallization period' brought up here, and accounts for
much of what has been discussed since.  
  I would amend the above hypothesis to read as follows:  Linguistic ability
(as an example of complex cognitively-learned behaviors, as opposed to things
like 3D visual perception) can only be brought about given a base containing
enough neural structure with a long period of highly dynamic change and
maturation and enough stimulation of the structure to organize it into 
function groups (ala neuronal nuclei).  This is a view of linguistic onset
and cognition in general that relies more on the developmental aspects of 
the brain than has been fashionable since the cognitive sciences became any
sort of a reality.  I do not believe that we will ever realize natural
language processing or any other sort of complex cognitive ability in 
artificial systems until we learn more about the development of the human
brain and take this information into account in our models.

Comments would be appreciated.


Mike Sellers
...!tektronix!sequent!mntgfx!msellers
Mentor Graphics Corp., EPAD

marty1@houdi.UUCP (12/04/87)

In article <1987Dec2.182753.622@mntgfx.mentor.com>,
msellers@mntgfx.mentor.com (Mike Sellers) writes:
> 
> ...  so I thought > I'd go ahead and throw in some neurological data....
> I'd appreciate comments, as I've not seen much in the way of cognition or 
> linguistics from a neurological point of view on the net.] ....

It's hard to begin to summarize a 200-line article that combines functional
and physiological observations and hypotheses in such a global way, but the
bottom line seems to be:

> Linguistic ability ... can only be brought about given a base containing
> enough neural structure with a long period of highly dynamic change and
> maturation and enough stimulation of the structure to organize it into 
> function groups (ala neuronal nuclei).

I don't have any facts to add to that.  I have generally been relying
on another person's experience in second-language teaching and in the
training of second language teachers, and on my own casual reading in
Science and Scientific American.  However, the above synthesis (and the
full discussion of which it is a summary) looks so good to me that I
don't want to let it drop without a ripple.  It makes a lot of sense.

But the later statement,

> ....  I do not believe that we will ever realize natural
> language processing or any other sort of complex cognitive ability in 
> artificial systems until we learn more about the development of the human
> brain and take this information into account in our models.

has to be read cautiously.  It means we need a good understanding of
the essential processes required to process language.  As has been
pointed out by others, it doesn't mean we should imitate structures
and techniques that are just one way of executing those processes.

M. B. Brilliant					Marty
AT&T-BL HO 3D-520	(201)-949-1858
Holmdel, NJ 07733	ihnp4!houdi!marty1

wcalvin@well.UUCP (William Calvin) (12/05/87)

Apropos cell death in brains, the old saw about losing 10,000 neurons every
day is now being challenged by the people that work on cerebral cortex; they
seem to think that there is little neuron loss there during most of postnatal
life.  Some subcortical areas like substantia nigra do lose 50% of cells by
age 70, while adjacent regions in midbrain may lose less than 2%.
  But there is a LOT of synapse death -- or, as I like to phrase it, withdraw
of axon collaterals, breaking synapses.  Synaptic density in neocortex
peaks at 8 months after birth (in humans; 2 months in monkey) -- and then
drops by 30-50% during childhood.  After puberty, the data gets too noisy
to interpret.  So there is a lot of opportunity for Darwinian editing of
randomly-made synaptic connection, achieving information storage by carving
(rather like photographic development removes unexposed silver grains).
	I review a variety of Darwinian selection stories in my piece
in the 5 November 1987 NATURE 330:33-34, entitled "The brain as a Darwin
Machine."
		William H. Calvin
		University of Washington NJ-15, Seattle WA 98195
		  206/328-1192  wcalvin@well.uucp

henry@utzoo.UUCP (Henry Spencer) (12/10/87)

> Apropos cell death in brains, the old saw about losing 10,000 neurons every
> day is now being challenged...

Also of note, the other old saw that new neurons are not formed after birth
in higher animals is now known to be untrue.  At least some higher animals
(some types of birds) do grow new neurons at times.  Last I heard, nobody
yet knows how common this phenomenon is.
-- 
Those who do not understand Unix are |  Henry Spencer @ U of Toronto Zoology
condemned to reinvent it, poorly.    | {allegra,ihnp4,decvax,utai}!utzoo!henry