From: Martin Striz (mstriz@gmail.com)
Date: Wed Apr 05 2006 - 14:06:54 MDT
On 4/4/06, Charles D Hixson <charleshixsn@earthlink.net> wrote:
> I'm sure you're right for an initial estimate. After things get working and
> start being understood, however, expect to be able to pare at least an order
> of magnitude off the estimate for the initial requirement.
>
> How much computing power you need is partially determined by how efficient
> your models are. If, e.g., you use a neural model of a clock, you will end
> up with both a clock that's rather inefficient, and an excessive number of
> neural connections. When it comes to other features, however, e.g., possibly
> pattern matching, our current estimates may be either right on the money, or
> even a bit low. Still, some neural connections are pretty clearly time delay
> loops, and those can be done much more efficiently that by neural simulation.
> This implies that there exist other features that once we understand, we will
> be able to simulate without replicating all the circuitry.
This is the most important post in this thread. Cognitive processes
are optimized for the kinds of computation that neural substrate does
well. Even if you knew how many computer-equivalent operations per
second a typical brain does, you wouldn't know how much computational
power computer chips require. Athlon and Intel chips perform
differently on different kinds of benchmarks. Neurons will have even
more significant deviations. At best you might be able to approximate
within an order of magnitude.
There are aspects of synaptic activity that simple circuits don't
capture, like receptor density, activiation length, that are involved
in probabilistic integration at the dendrite. I would lean to the
side of more bits rather than less to describe a synapse, divided by
redundancy.
A reasonable model is Dani's: 10^15 to 10^18 computer-equivalent
flops. But even if you have that much processing power, you have to
know what to do with it.
Martin
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:56 MDT