[ExI] ai class at stanford

[Adrian Tymes]

On Mon, Aug 29, 2011 at 12:16 AM, G. Livick <glivick at sbcglobal.net> wrote:
> Hardly.  The scientific method places the onus on the claimant.  Otherwise,
> we'd have to accept everything imaginable as possible, since "nothing is
> impossible."

Actually, no.  There are some things that have been proven impossible.  For
instance, accelerating (in the usual sense) from below the speed of light to
above it, for any entity that has mass.  The class of things that we merely
don't yet know how to do is different.

> One will not find such claims about the workings of the brain, let alone
> mind, in the scientific literature.

Because such speculation does not belong in the scientific literature.  Said
literature is for reports of results after the work has been
performed.  Instead,
look for project descriptions, that - if successful - will ultimately result in
reports in scientific literature.  Look up the Blue Brain Project, for
example: it
is explicitly trying to simulate an entire mammalian brain.

> The layman's knowledge of the state of
> the art of things does not open for him the awe of the researcher over how
> little is known.  As a result, an unfounded optimism is not unusual among
> interested observers.

This is true, and we agree that the AI course shall inform us of the state of
the art better.

> As for the number of "AI Techniques" we already have
> being nearly sufficient, after a little more tweaking, to knit into a grand
> unified solution that will cover all bases: the people taking the Stanford
> AI course should emerge from it able to assess that statement directly.  My
> fiddling with this stuff over the years has me thinking quite the contrary,
> though; the field is in it's infancy as I see it.

As I said, the pieces need improvement on their own.  It's a lot more
tweaking, not a little.

> Technology is developed from previous technology in incremental fashion.  It
> never emerges out of whole cloth.

This is not always true.  There are many documented examples of revolutionary,
rather than evolutionary, improvements.

But that is irrelevant here.  There appears to exist a path, using only
incremental improvements, from where we are today to emulating human brains,
and from there to translating an existing, living brain into such an emulation.

> True enough.  But without the man making it work, there would have been no
> magic.  Dorothy's observation of it did not cause the magic to present
> itself to her.

Agreed.  And this course is about creating said man.

> It's not just impractically expensive to replicate the actions of a neuron
> in software today, it's impossible.  We don't have the whole picture yet;
> I'd guess, at best, only 10%, of which 9% will be shown in time to be
> inaccurate.

I said it's impractically expensive to replicate an entire brain, not a single
neuron.

As to emulating a neuron - a quick search on "neuron simulation" seems to
provide many examples of replicating the actions of a neuron in software
today.  But for sake of argument, let us assume they are all imperfect.

Are you saying that it is physically impossible - not just "we don't have the
information today", but "we can never obtain this information" - to eventually
discover how neurons work, in 100% detail?

This seems to be the core of your argument, and an extraordinary claim.
What evidence do you have that it is impossible to uncover the missing
details?

Note that we have 100% detail about certain other physical phenomena.
For instance, how electricity travels through a wire made of thickness X
by length Y of material Z, or how air under given atmospheric conditions
travels over a certain surface.  Neither is mere complexity a blocker:
there are many programs that have taken the air flow model, and applied
it to the airflow over an entire airplane under a series of conditions, for
instance to find where on the plane a flight will induce the most stress -
and an entire 747, say, would seem to be at least as complex as a
typical biological cell.  If it is impossible to get this level of detail about
neurons, what is it about neurons that makes this impossible?

If it is not impossible to uncover said details, then it is not impossible to
emulate them in software once they are uncovered.  Once we can emulate
one - including a complete model of its synapses - then we can emulate
two, including any synapses they share if the original model correctly
emulated the one neuron's synapses (and if it is confirmed that they only
interact via synapses - if there is any other method, we will need to uncover
and emulate that too).  Once we have two - again, including their method of
interaction - then we can add a third, and a fourth, and eventually an entire
brain's worth.