The page this refers to

My response


Date: Thu, 8 Jul 1999 03:37:39 -0500 (CDT)
To: yeh_harlequin.co.uk
Subject: Carla Shatz
From: "James A. Bednar" 



Mr. Harpaz,

I am not Carla Shatz, nor do I work with her, but I do happen to be
undertaking a Ph.D. in a topic closely related to her research.  I am
quite happy to reply on her behalf to your concerns about her public
comments, which you have posted at:

   http://human-brain.org/neuropages.html 

First, I agree that her comments to the public do not neccessarily
convey the precise, subtle nuances of everything that is going on
during neural development, and I do not think that she would claim
that they do.  Then again, I think that such a disclaimer is valid for
any public comment by any scientist at any time.  Only very simple
concepts can be communicated directly to the non-specialist; for
anything else one must revert to metaphor or else remain entirely
silent.  Since she is explicitly using vague metaphors like "nerve
cells are kind of like telephones", I don't think many people
listening will really think they are hearing every bit of the story.

Second, I will even grant that some of her comments end up being
misleading when interpreted in a fairly reasonable way.  However, I
frankly doubt that anyone else could have summarized that topic for
the nonspecialist any more accurately than she did, and I do not
believe that any of her statements are misleading in context.  

My specific replies are below.

|  1. Carla Shatz about connectivity in the brain
|                                   
|  [1.1] On 17Apr97, Carla Shatz took part in a white-house conference on
|  childhood development an learning[23Feb99: this page disappeared]. A
|  more organized version of her remarks is here, and I will use the
|  latter. This talk is a good example of neuroscientists confuse the
|  rest of the public about the precision of connectivity in the brain,
|  even though they know better.         
|                                   
|  [1.2] The main point of Dr. Shatz talk is that experience after birth
|  is important in the wiring of the brain, and I have no problem with
|  this point. The problem is what Dr Shatz says about the precision of
|  connections.                  

Actually, one crucial point she mentions is that experience is
important both before *and* after birth.  For a review of the relevant
experiments see e.g. wong:arn99, in the references at end of this
message.

|  [1.3] In the second paragraph of the 'The problem' section, Dr Shatz
|  says: "So, the brain contains well over 1000 trillions connections and
|  none of them are random!" In the following paragraph, she expands a
|  little.  Discussing visual and auditorial input, she says:"And then,
|  once the growing axons reach their targets, they must select, from the
|  millions possible neurons, just the right few with which to form
|  synapses."
|
|  [1.4] First we should note that Dr. Shatz don't actually tells us what
|  'not random' or 'right few' means. The natural interpretation, which
|  non-neuroscientists will use, is that there is some plan for the
|  connections, and each neuron select the right target neurons in
|  accordance with this plan. Thus, for a non-neuroscientist, what
|  Dr. Shatz says means that there is a plan which specifies exactly all
|  the connections in the brain ("none of them are random").

I agree that some people might come to such an interpretation based on
what you quoted, particularly non-specialists.  I also agree that the
definition of "right few" and "not random" are crucial.  

However, I do not believe that many people will come to interpretation
[1.4] based on her talk, since she quite explicitly states that:

  ...there are just not enough genes in the genome (only about 100,000)
  to account for the incredible precision of connectivity present in
  the adult brain (> 1000 trillion connections)...

So she quite clearly and without ambiguity states that there is no
fully-specific genetic plan of connections, nor could there be one.
And that's the key question of her research: how can things not be
entirely random if there is no specific blueprint?

|  [1.5] That is simply false. The main structure for thinking in the
|  human brain is the cerebral cortex, and in it specific connections are
|  almost totally stochastic. There are tendencies in coarse resolution
|  (> 1mm), but at the level of individual connections each brain is
|  completely different from each other brain (That is true for all
|  mammals). In the cortex, neurons don't select the 'right few' neurons
|  to synapse with: they synapse stochastically with some neurons that
|  happen to be around the point their processes 'decide' to start to
|  make connections.             

I agree that interpretation [1.4] is false.  However, she herself
states that the answer to the key question is that:

  ... An elegant solution is to 'hard wire' the trunk lines with
  specific molecular guidance clues, but then flip the switch ON early
  and let neural function make the final decisions.

Thus she would not contradict your assertion that each particular
connection is not precisely identical between individuals.

However, as I argue below, it does not follow from such arguments that
neurons do not select the "right few", nor does the issue of the
initial stochasticity have any bearing on whether the right few
neurons are eventually chosen.  As she states, the key issue is that
"neural function" is what does the choosing, not some
genetically-defined, preexisting label.

|  [1.6] Dr. Shatz knows this better than I do, as she has been
|  researching the connectivity of the brain for years, and never found
|  individual connections in the cortex of any mammalian brain which are
|  the same between individuals. Nor did any other the other researchers
|  in the field.                         

I am certain that Dr. Shatz (and probably every other neuroscientist
alive today) would also disagree with interpretation [1.4].  

|  [1.7] What about the examples that Dr. Shatz quotes, with the neurons
|  that select the 'right few'? These example are wrong on two accounts:
|                                   
|  They are not accurate.        
|       Neurons which selects exactly which
|       neurons to contact (i.e. they have the same connections in all
|       healthy individuals of the same species) haven't been seen
|       anywhere in any mamalian brain. 

Here you are interpreting "right few" as some sort of preexisting
genetic label that determines a neurons purpose for all time, which is
a reasonable but incorrect interpretation of her remark in context.
What she means by "right few" is instead functional, i.e. the neuron
that has a certain role in the system.  As I said above, she states
this explicitly when she says that "neural function" is what
determines the connectivity.

|       The most accurate connectivity is
|       in cases when small populations of neurons are directed to other
|       small populations (e.g.  neurons from a specific whisker to a
|       specific barrel in the rat), but even in this case the
|       connections of each neuron are not completely specified ([As far
|       as I know, this is true not only in mammalians, but in all
|       vertebrates, and for most invertebrates that have a brain worth
|       talking about.]).        

Her concept of "trunk lines" explicitly addresses this issue, as I
quoted above; she does not argue that individual neurons in each
whisker barrel have predetermined roles.

|  Even that level of specificity is seen only outside the cerebral cortex 
|       The examples that Dr. Shatz brings are pathway leading to the
|       cortex. These are important too, but when Dr. Shatz discusses the
|       '1000 trillions connections' she clearly means the
|       cortex. Outside the cortex, the only structure that contains that
|       many connection is the cerebellum, which has a major role only in
|       the learning and performing of accurate movements. The number of
|       connections in the rest of the brain is far smaller than that.

I agree that she is here referring to the cortex.  However, her
argument requires only broad specificity, fully consistent with that
seen in the cortex.

|  [1.8] In 'The solution - not like a computer' Dr. Shatz describes how
|  she thinks the precise connections are made: "The brain actually lays
|  down a basic framework of circuits - the trunk lines - according to
|  strict diagrams set by genetic blueprint. Then, way before the adult
|  precise circuits are formed, the "switch" is turned on: Brain function
|  itself completes the wiring process by running test patterns on the
|  circuits and selecting correct connections and eliminating errors."
|                                   
|  [1.9] The first sentence of this quote is a half-truth: while the
|  brain does have a design, this design is in coarse structures. The
|  exact connectivity of neurons is not defined at any stage of the
|  development of the mammalian brain. If we remove the word 'strict'
|  from this sentence, it is much closer to the truth.

I agree that "strict" would be better left out; I would prefer
"approximate".  But that's a very minor point, given that she is
basically making a comparison between the large scale and small scale
connectivity.  I think it is quite reasonable to claim that the large
scale has a "strict diagram" relative to the small scale.

|  [1.10] The second sentence is blatant nonsense: How can any part of
|  the brain 'run test patterns' and 'select correct patterns' before it
|  itself is connected properly? Dr. Shatz implicitly uses the analogy of
|  electrical engineer running tests on an electrical circuit, but in the
|  brain we don't have an engineer, and we don't have any mechanism that
|  can 'select correct connections', because this mechanism itself would
|  have to based on 'the correct connections'
|  [Note that I am not saying that there are no activity-dependent
|  changes, or that these are not essential for normal development. It is
|  just that these cannot be described as 'running test patterns' and
|  'selecting correct connections'.].

Ah, here is the big confusion.  Actually, running test patterns is one
reasonable description what the brain seems to be doing, though I've
never heard it described that way before.  Before eye-opening, the
spontaneous activity in the eye (i.e. the retinal waves) does indeed
seem to act like a test pattern, i.e. like a calibrated set of images
with which to evaluate a system (wong:arn99).

So your key criticism must be about the "engineer" metaphor.  Here
again "correct" must be interpreted in terms of function, NOT in terms
of some genetic label that the neuron was born with.  That is, both
the connections and the neurons are being sorted out at the SAME time;
being the correct neuron depends upon having the correct connections
and vice versa.  The rules to sort this out without any need for an
external engineer are quite simple; e.g. see the Hebbian models of
bednar:cns97 and haith:cns97, as well as keesing:nips91.  In these
models "correct patterns" of connections are indeed selected, but they
are correct only because of the way that they work, not because of
some predefined genetic label.

|  [1.11] The second sentence also enforces the notion of a plan: It does
|  not make sense to talk about 'selecting correct connections' and
|  'eliminating errors', unless there is some plan so 'correct
|  connections' are connections according to the plan, and errors are
|  not. Thus Dr. Shatz does not leave any doubt in the non-neuroscientist
|  reader/listener that she thinks there is a plan for the connections in
|  the brain, which specifies all the connections.

Actually, having a specific plan is just one way to have "correct
connections"; all that she means by "correct" is that it works
correctly, not that it matches some blueprint.  Consider an simple
analogy: 

  Let's say I'm outside one day, it's hot, and I'm thirsty.  I start
  walking to find a store to get something to drink.  I happen to see
  a little kid selling lemonade, so instead of going to a store, I
  give some money to the kid, get some lemonade, and go on my way.

This was a successful ("correct") transaction -- I wanted something
very specific, and it happened.  How did it work?  Well, I made a very
specific, correct, and *non-random* connection with the kid selling
lemonade.  If that kid had been selling air freshener I would have
just kept on going, so it was clearly not a random act.  But on the
other hand, that kid was not destined from birth to sell me lemonade,
nor is there some master plan filed with the city that that kid should
provide the lemonade that I need.  I was even planning to go to a
store instead.  Yet I was able to tell just from the kid's actions
(i.e. standing on the street selling lemonade) that he was the correct
connection for me, and I was totally correct about that.  Of all the
people in the city, only the "right few" would have been willing to
sell me a drink, and I found one, and I could have found more if I had
walking.  If instead when I got up to him, I saw him start pouring
that lemonade into a car radiator, I would realize by his actions that
he was *not* the right connection for me (i.e. he was really holding
antifreeze and not lemonade) and I would move on to some other store.

So the key principles should be clear -- selecting the "correct"
"right few" targets in a "non-random" way is totally compatible with
the actual targets being stochastically-determined in some sense.
"Errors" are only known by the actions, e.g. when the target fires and
you don't, in Hebbian learning.  "Correct" is also known only by the
actions, in the same way.  No engineer is required, and no blueprint,
yet specific and correct.

|  [1.12] Why does Dr. Shatz tell her listeners/readers that there is a
|  plan for the connections in the human brain, when there clearly isn't?
|  First we should note that she does not actually say it: It is just
|  strongly implied by the text. Hence one possibility that Dr. Shatz
|  does not intend to imply that there is a plan for the connections.

It is true that she does not actually say that, but I disagree that
she implies it, since she quite explicitly says that there is not
enough information in the genome to specifically encode each
connection.  Thus at worst she implies a paradox, which is actually
very close to the truth.

|  [1.13] That is quite difficult to reconcile with the terms she uses (
|  "select .. the right few", "select correct connections", "eliminate
|  errors"), but maybe in neuroscientists's jargon these have different
|  meaning. In this case, that text (and other neuroscience texts) is
|  very misleading for other people, which read it using the normal
|  definitions of these terms, and hence take it as implicitly saying
|  that there is a plan.                 

I think that indeed many people would hear "select .. the right few",
etc. and think that there is a strict plan, but when they hear that
there is not enough space in the genome to encode a strict plan, then
they would either:

  a) be confused (which is ok, because it is a confusing issue which we
     are only now starting to figure out!), or 

  b) get the right idea, which is that it all works out somehow, it 
     seems to depend on internally-generated activity, it seems to
     depend on the environment, it's not totally obvious how it all
     happens, but we are starting to figure it out.

|  [1.14] The other possibility is that Dr. Shatz does believe that there
|  is a precise plan for the connections in the brain, including the
|  cortex. As I wrote above, this is simply false, because different
|  brains have different connectivity, and Dr. Shatz knows that. In this
|  case this is an example of a person believing two contradicting facts
|  in the same time.             

There is actually not any reason to conclude that she is believing two
contradicting facts at the same time, since they do not in fact
contradict.  The key issue is that "correct" does not in fact mean
"genetically-determined", which I believe she gets across quite well
in her short talk, despite the impression given by the quotes taken
out of context.

|  [1.15] In either case, an underlying reason for the mistake that
|  Dr. Shatz (and other neuroscientists) does is the pressure from
|  outside the field of neuroscience. This comes both from cognitive
|  scientists, which advance theories that are incompatible with
|  stochastic connectivity, and from the rest of the public, which
|  prefers more 'positive' statements. Thus a statement like "the brain
|  contains well over 1000 trillions connections and most of these are
|  stochastic", even though much more accurate than the statement
|  Dr. Shatz makes, would be received much less favourably both by the
|  cognitive scientists and the rest of the public.

I think her talk was deliberately designed to get across her point to
the nonspecialist, and no assumption of political intrigue or false
optimism is required.  In fact I frankly doubt that the entire field
of cognitive science has much influence on the field of neuroscience.
I consider it regrettable that they overlap so little and would indeed
rather see an increase in overlap to be desirable.  I do not believe
that any of Dr. Shatz's comments contradicted the idea that there is
stochasticity involved.  Moreover I doubt that the general public
would be any less alarmed to think that their entire brains are
genetically determined in a specific blueprint than to think that
there is there is disorder involved at some level.

|  Dr. Shatz response (If I get any)

You are welcome and encouraged to post this as my response on her behalf.

Cordially,

James A. Bednar
Department of Computer Sciences                         
The University of Texas at Austin


(Yes, that is correct; I am nominally in the Computer Science Department.)
-------------------------------------------------------------------------------


@Proceedings{cns97,
  editor       = "J. M. Bower",
  title        = "Computational Neuroscience: {T}rends in Research,
                  1998",
  booktitle    = "Computational Neuroscience: {T}rends in Research,
                  1998",
  publisher    = PLEN,
  address      = PLEN-ADDR,
  year         = 1998,
  annote       = "This book is the Proceedings of the Sixth Annual
                  Computation and Neural Systems Conference (CNS*97)",
}

@InProceedings{bednar:cns97,
  author       = "James A. Bednar and Risto Miikkulainen",
  title        = "Pattern-Generator-Driven Development in
                  Self-Organizing Models",
  crossref     = "cns97",
  pages        = "317--323",
  url          = "http://www.cs.utexas.edu/users/nn/pages/publications/abstracts.html#bednar.cns97.ps.Z",
}

@InProceedings{haith:cns97,
  author       = "Gary Lawrence Haith",
  title        = "A Computational Model of Retinogeniculate Development",
  crossref     = "cns97",
  url          = "ftp://white.stanford.edu/users/haith/cns97_submitted.ps.gz",
}

@InProceedings{keesing:nips91,
  author       = "Ron Keesing and David G. Stork and Carla J. Shatz",
  title        = "Retinogeniculate Development: {T}he Role of
                  Competition and Correlated Retinal Activity",
  pages        = "91--97",
  crossref     = "nips91",
}

@Article{wong:arn99,
  author       = "Rachel O. L. Wong",
  title        = "Retinal Waves and Visual System Development",
  journal      = ARN,
  year         = 1999,
  volume       = 22,
  pages        = "29--47",
  bibdate      = "Fri Jun 11 17:28:55 CDT 1999",
  bibauthor    = "jbednar",
}