From: Richard Loosemore (rpwl@lightlink.com)
Date: Fri Jun 02 2006 - 07:47:13 MDT
If I may: this seems to be an example of what has come to be a standard
calculation of <When Gate-Count Will Equal Neuron-Count>. People come
up with different numbers, of course, but lots of people do the same
calculation.
Now, you point out that this is only some kind of upper bound, and that
it may not be as important as (e.g.) architecture ...... but to my mind
this kind of calculation is a *complete* distraction, telling us almost
nothing but making us think that it means something.
In my estimate, the implementation in neural hardware is not necessary,
and I believe that we have the hardware to do it now. In fact, I think
we have had that hardware for about a decade. That is a rough estimate
based on my understanding of human cognitive architecture, and my take
on what the design of the first successful AGI will be like.
Richard Loosemore
Keith Henson wrote:
>
> [Reposted from another list with permission. Nothing new, but an
> indication that the local topics are being discussed elsewhere -- Keith
> Henson]
>
>> Date: 31 May 2006 10:35:31 -0800
>> From: dan miller <danbmil99@yahoo.com>
>> Subject: Re: Moore's Law and AI (Real or Artificial Intelligence): was
>>
>> ( the following offered as a simulus for discussion and debate; I'm not
>> claiming it's scientifically rigorous )
>>
>> I think it's possible to put forward a somewhat reasonable estimate of
>> computing power necessary to roughly equal human-level intelligence.
>> If we
>> look at a typical insect, which has on the order of 20,000 - 200,000
>> neurons
>> (I know, not all neurons are created equal, but this is
>> back-of-the-envelope) -- we can ask ourselves, how does this setup
>> compare,
>> in terms of "intelligence", to a silicon-based machine that has similar
>> capabilities?
>>
>> [caution: arm-waving begins]
>>
>> I conjecture that a typical Darpa GC vehicle represents a similar
>> level of
>> complexity in terms of its ability to sense, react, and (to a degree)
>> plan
>> its behavior within its environment. Clearly there are many differences,
>> but I'm pretty sure it's within an order of magintude one way or the
>> other.
>>
>> The GDC vehicles were designed as one-off prototypes, so the
>> technology used
>> was not highly optimized for low cost, power consumption, etc. CMU and
>> Stanford both used about half a dozen powerful PC's each; but it's
>> obvious
>> to me that optimizations, including special-purpose chips or FPGA's,
>> could
>> reduce that requirement by at least an order of magnitude.
>>
>> So conservatively, a present-day-class, 2+ ghz Pentium-based computer is
>> capable of emulating the functional capabilities of something like an
>> ant.
>>
>> So 2G mips ~== 20K neurons; one neuron = 100,000 mips
>>
>> Humans have on the order of 10^11 neurons; 10^11 * 100,000 = 10^16 mips
>>
>> After sketching this out, I looked up Hans Moravec's estimate, which is
>> 10^14. I guess he's planning to write his neuron simulators in assembly
>> code.
>>
>> My engineer's gut tells me this estimate is an upper limit, and that
>> appropriate special-purpose hardware would enable the right sort of
>> computational horsepower attainable at reasonable cost within 10 to 15
>> years.
>>
>> It's interesting to note that if the typical guesses are correct,
>> Google is
>> just about at this level of computational ability.
>>
>> None of this is meant to suggest that the architecture isn't more
>> important
>> than the gate count; but it's nice to have some likely upper bounds on
>> what
>> kind of performance you might need to get to serious AI.
>>
>> - -dbm
>
>
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:56 MDT