J. Andrew Rogers wrote:
On Dec 18, 2008, at 10:09 PM, Colin Hales wrote:
I think I covered this in a post a while back but FYI... I am a
little 'left-field' in the AGI circuit in that my approach involves
literal replication of the electromagnetic field structure of brain
material. This is in contrast to a computational model of the
electromagnetic field structure.
Here is a silly question:
If you can specify it well enough to implement the desired result in
hardware, why can't you implement it in software? It is equivalent,
after all.
And if you can't specify the dynamic well enough to implement it
virtually, why would there be any reason at all to believe that it
will do anything interesting?
The hallmark of a viable AGI theory/design is that you can explain why
it *must* work in sufficient detail to be implementable in any medium.
J. Andrew Rogers
"/If you can specify it well enough to implement the desired result in
hardware, why can't you implement it in software? It is equivalent,
after all./ "
The answer to this is that you /can /implement it in software. But you
won't do that because the result is not an AGI, but an actor with a
script. I actually started AGI believing that software would do it. When
I got into the details of the issue of qualia (their role and origins) I
found that software alone would not do the trick.
If an AGI is to be human equivalent, it must be able to do what humans
do. One of those behaviours is science. Getting the 'logical dynamics'
of software to cohere with a 'law of nature' is, I believe, impossible
for software alone, because the software model of the dynamics cannot
converge on an externally located and intrinsically unknown (there is no
model!) knowledge. How a software model of a modeller of the
intrinsically unknown (a scientist) can work is something I have had to
grapple with. In the end I had to admit that software seemed less
plausible than actually implementing the full physics of brain material.
Hence my EM approach.
The simplest way to get to the position I inhabit is to consider that
the electromagnetic field has access to more information (about the
world outside the agent) than that available through peripheral nerve
signaling. It's the additional information that is thrown away with a
/model/ of the electromagnetic field. It's replaced with the arbitrary
and logically irrelevant electromagnetic fields of the computer
substrate (basically noise). The spatially expressed EM field inherits
(dynamics is altered by) information from the external world directly
via space..
The EM fields play a very important role in the dynamics, adaptation and
regulation of neural activity and none of it is captured with existing
neural models - as it acts /orthogonally/ , coupling neurons spatially
via EM events, not dendrite/axon routes. It's outside the neurons in the
spaces in between. It's the reason cortex is layered and columnar.
Cortex is 50% astrocytes by volume. They are all charged up to -80mV.
and are intimately involved in brain dynamics. Because the boundary of
the cells and space is as much an information source as all the
peripheral nerve 'boundaries' (the surface of your body), and the
boundary is literally electromagnetism (there's nothing else there!),
you can't model it for the same reason you can't model the peripheral
nerve signals (you have to have the EM fields for the same reason that
you need to have a retina or camera)... by extrapolation everything else
follows.
The EM coupling effects are the subject of my PhD and will be out in
detail ASAP. Bits of it will be published - It's been a real trial to
get the work into print. I tried to get a publication into the AGI
conference but ran out of time.
The original hodgkin-huxley model (upon which all modern neural
modelling is based) threw out (or at least modelled out) the EM field.
If you look in the original 1952 papers you'll see there are batteries,
non-linear, time-varying resistors and,,,, ignored and off to one side
all by itself, waiting patiently .... a little capacitor. That little
capacitor hides the entire EM field spatial behaviour. If you drew the
model properly all the components in the model actually span the
dielectric of the capacitor between its little plates. The capacitor is
actually linked to lots of other capacitors in a large 3-dimensional
mesh. You can't delete (via a model) the capacitors because their
dynamics is controlled (very very lightly but significantly) by the
external world.
So...My approach puts the fully spatially detailed EM field back into
the model. The little HH capacitor turns into an entire new complex
model operating orthogonally to the rest of the circuit. That capacitor
radically changes the real model of brain material. There is spatial
coupling to other neurons that happens using the field that has been
averaged out and confined inside the dielectric of the capacitor, It's
been waiting for someone to find it for 50 years.
Functionally, the key behaviour I use to test my approach is "scientific
behaviour". If you sacrifice the full EM field, an AGI would provably be
unable to enact scientific behaviour because the AGI brain dynamics
would be forced to operate /without the dynamics of the EM field/, which
is literally connected to the distal natural world (forming a new I/O
stream). The link to the distal natural world is critically involved in
'scientific observation'. You can't simulate it because it's what you
are actually there to gain access to. A scientist does not already know
what it 'out there' - an AGI scientist needs what human scientist has
in order that the AGI do science as well as a human. Scientific
behaviour easily extends to normal problem solving behaviour of the kind
humans have. Hence 'general intelligence'.
One way of understanding why you can't use software is that you'd have
to know everything already (all science must be complete, including
already knowing the world surrounding you!). We humans do not know
everything already - because science is possible! I rather like the
finality of that logic. :-). The whole problem of AGI is not how to be
smart - but how to be wrong.then overcome that ignorance. If you write
software you are completely logically fixing /-forever-/ how ignorance
shall be handled. "/All novelty shall be handled thus/". As a result the
external natureal world's real rules could become inaccessible to your
software AGI.
In adopting 'the scientist' as my AGI benchmark, I find I must
replicate, in my AGI, all critically involved human physics. Hence the
need for the new chips. It may be in time that certain portions of the
chips are suited to software replication. However - until I have a full
working (very dumb - but real) EM AGI running, I won't be able to know
which bits of the dynamics can be abstracted away without impoverishing
the intelligence of the AGI. I don't have the confidence to make that
decision now. The physics has to be done properly so that decisions can
be made based on solid knowledge. We currently do not have that knowledge.
So in your original ""/If you can specify it well enough to implement
the desired result in hardware, why can't you implement it in software?
It is equivalent, after all./ " .... I'd modify it.
(a) You can't 'specify it' completely in the case of a scientist AGI.
because the resultant behaviour is unknown and unique.
(b) _Yes - Software can be equivalent._ But in the case of an AGI
capable of science, /software is not sufficient/ because you'd have to
know everything already (pre-program all ignorance, if you like!).
That's probably not a satisfactory answer, but it's all I can do in a
few sentences.
cheers
Colin Hales
-------------------------------------------
agi
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/
Modify Your Subscription:
https://www.listbox.com/member/?member_id=8660244&id_secret=123753653-47f84b
Powered by Listbox: http://www.listbox.com
