From: AD.COM - Marcos Guillen (marcos@ad.com)
Date: Tue Sep 16 2003 - 16:31:21 MDT
>I saw the same talk as Ben, and his post pretty much sums it up. The
>current model of the brain this group has is rather simplistic, to say
>the least. On the other hand, the architecture of their system seems
>to allow fairly rapid additions to the model.
>
>
>From: Ben Goertzel [mailto:ben@goertzel.org]
>> I just saw a talk by one of the leaders of this enterprise,
>> at the Accelerating Change 2003 conference.
>>
>> It's an interesting effort, but they seem to have no
>> particular insight into the operation of the human brain ...
Yea, I agree. At Artificial Development, we have no particular insight into
the operation of the human brain, and we are rather simplistic folks.
Further more, we are not that smart, either. Nothing to see here, keep on
moving :-)
Anyway, I would like to thank you both, Ramez and Ben, for attending our
presentation, it was a pleasure talking for such a qualified audience. It is
great to be able to exchange ideas with people like you.
We are computer engineers, no neurologist, and we are aware that our
immediate goals may not seem that interesting from a theorist perspective.
Our goal is to build a system capable of algorithmically operating over
billions of neuron and trillions of connections, in real time. Sounds fun to
you? Well, for us it is, and it's quit a challenge: CCortex data matrix is a
number of times bigger than Google.
You seem to be concern with what you perceive as a 'simplistic' approach, in
what specific operations do we perform, and what parameters we are
considering. Well, so we do. We don't think that there is enough knowledge
out there to define the precise operations we must necessarily perform to
correctly update our artificial brain. We do not have any special,
'particular insight' into the human brain, no mystic, sudden revelation here.
In fact, the only thing we are really sure about is that neither we, neither
other competing groups know enough at this point; and those are precisely
the foundations of our project: A versatile, flexible, powerful, easy to
adapt, pure software implementations.
By design, we can modify the central algorithm governing the update over the
weekend, to accommodate different theories. We keep individual memory space
for each neuron to accommodate a number of fields of updatable data
pertaining to the neuron body, plus vector, positioning and weigh data
fields for and average 1,000 connections per neuron. What fields? Those we
deem necessary in each run. What operations? Any operation we may need. Most
operation has low computing requirements, while the rest are being
approximated or pre-calculated and stored on a database. After all, after a
few cycles, the bulk of the possible operations to perform in our data space
has been already calculated, and is susceptible of being sampled, stored and
reused with minimum error margins.
Finally, keep in mind that, also by design, we are a privately held company,
and we are not aiming for the sort term. We do not depend of outside
financing, and we do not need to spend that much time with things like
customers, academic peer approval, or even (mostly nonexistent) competitors.
And yes, we are here to stay, but don't expect our 1 week old cluster to be
ready for prime time any time soon: this are just the first steps on a long
term project, and it's no time yet for Eliezer to start freaking out :-)
Best Regards,
Marcos Guillen
President and CEO
Artificial Development, Inc.
WWW.AD.COM
Best Regards,
Marcos Guillen
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:42 MDT