Joshua Fox wrote:
Greetings, I am new to the list. I hope that the following question adds
something of value.
Estimates for the total processing speed of intelligence in the human
brain are often used as crude guides to understanding the timeline
towards human-equivalent intelligence.
Would someone venture to guesstimate -- even within a couple of orders
of magnitude -- the total processing speed of higher order cognitive
functions, in contrast to lower-order functions like sensing and
actuation. (Use any definition of "higher" and "lower" order which seems
reasonable to you.)
I appreciate the problems with estimating human-equivalent intelligence
based on raw speed, and I recognize that tightly integrated lower-order
functionality may be essential to full general intelligence.
Nonetheless, it would be fascinating to learn, e.g., that the "core" of
human intelligence use only 1% of the total power estimated for the
brain. That would suggest that /if/ lower order functions can be
"outsourced" to the many projects now working on them, and offloaded at
runtime to remote systems, then human-order raw power may be closer than
we thought.
Joshua
Joshua,
I recently addressed a similar issue on the SL4 list, so here is an
expanded version of my calculation for what I think is involved in
higher order processing. My thoughts were geared towards estimating
when the hardware would be available. (Answer: yesterday.)
1) Quick Introduction
The basis for these calculations is the idea that the human cognitive
system does all of its real work by keeping a set of elements
simultaneously active and allowing them to constrain one another.
Simple enough idea. Basis of neural nets, actors, etc.
Then, starting with this idea, I use the fact that the brain is
organized into cortical columns, and I would (cautiously) hypothesize
that these could be implementing a grid of cells on which these elements
can live, when they are active. This allows us to start talking about
possible numbers for the simultaneously active elements and their
operating timescale.
Finally, notice that a good chunk of the cortical column real estate is
probably devoted to visual processing. Now, some of this would not just
be doing data driven processing (which would come under the heading of
"peripheral" work, which we want to keep out of the calculation) but
interactive processing that includes top-down constraints. Difficult to
say how much of this visual processing really counts as higher order
thought, but my guess would be that some fraction of it is not.
2) The Calculation Itself
Approximate number of cortical columns: 1,000,000. If each of these is
hosting a single concept, but they are providing a facility for moving
the concept from one column to the next in real time, to allow concepts
to make transient connections to near neighbors, then most of them may
be just available for liquidity purposes (imagine a chinese puzzle on a
large scale... more empty blocks means more potential for the blocks to
move around, and hence greater liquidity). So, number of simultaneously
active processes will be much less than 1,000,000.
My use of the cortical coumn idea is really just meant as an upper
bound: I am not committed to this interpretation of what the columns
are doing.
Second datum to use: the sensorium (the sum total of what is actively
involved in our current representation of the state of the world and the
content of our abstract thoughts) is likely to contain much less than
1,000,000 simultaneously active concepts. Why? Mostly because the
contents of a good sized encyclopaedia would involve less than a million
concepts, and we barely have enough words in our language for that many
distinct, nameable concepts. It is hard to believe that we keep
anything like that many concepts active at once.
Using the above two factors, we could hazard a guess at perhaps as few
as 10,000 simultaneously active high-level concepts, not a million. My
gut feeling is that this is a conservative estimate (i.e. too high).
Further suppose that the function of concepts, when active, is to engage
in relatively simple interactions with neighbors in order to carry out
multiple simultaneous relaxation along several dimensions. When the
concepts are not active they have to go through different sorts of
calculations (debriefing after an episode of being used), and when they
are being activated they have to (effectively) travel from their home
column to where they are needed. Considering these "other" computations
together we notice that the cortical column may implement multiple
functions that do not need to be simultaneously active.
Now, all of the above functions are consistent with the complexity and
layout of the columns. Notice that what is actually being computed is
relatively simple, but because of the nature of the column wiring the
functions take a good deal of wiring to implement the functions ... so
the columns look computationally demanding but when implemented in
silicon the functionality is not nearly as difficult.
Finally, when implementing these 10,000 processes in silicon, take
account of the relative clock speeds and you can probably simulate 100
to 1000 processes simultaneously, if you use FPGA hardware (like one of
the Celoxica boards that Hugo de Garis is making such good use of).
The exact amount of hardware required depends on the complexity of the
function computed by each element, and on the bandwidth requirements.
But assuming that one FPGA board can reliably maintain only 100 to 1000
elements, that implies a computational requirement of between 10 and 100
desktop machines with one $6,000 FPGA card in each one. (Obviously that
is just the cognitive core: you'd need peripherals as well).
Assuming 100 machines rather than 10 (i.e. erring on the conservative
side again), and another fifty equivalent machines for peripherals, that
would be an approximate AGI cost of roughly $1 million.
3) Timeline
When will this become available? If you have a million dollars to
spare, it already is.
How long has this been available? If money was no object, it was
available a couple of decades ago.
4) Postscript
I really think this is too conservative. I don't believe there are as
many as 10,000 concepts involved, but only a few thousand. A few
hundred is too small, but between one and a few thousand seems about right.
What this implies is that the big obstacle is not how much hardware you
have got, but how you use it. My personal opinion is that AGI could
have been achieved a couple of decades ago, and that what stopped it
from happening was a lack of understanding of the nature of the software
problem. That lack of understanding persists.
Or, as Bananarama might have said: It Ain't What You Got, Its The Way
That You Do It.
Richard Loosemore.
-------
To unsubscribe, change your address, or temporarily deactivate your subscription,
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
