From: Dani Eder (danielravennest@yahoo.com)
Date: Thu Jan 12 2006 - 08:14:47 MST
> On my machine, a 3GHz workstation, im able to run a
> feedforward network at
> about 150.000 operations /second WITH
> training(backprop) .. take training
> out of the equation and we may, lets shoot high,
> land on 1 million 'touched'
> neurons/second .. now from 10^6 -> 10^14 .. that's
> one hell of a big
> number!!
>
You are losing a lot of orders of magnitude by
emulating
a neural network on a Von Neuman architecture. Taking
the
reverse, our brains, which are neural nets, can
emulate
a VN machine at about 1 op/sec (doing mathematical
calculations).
We are much faster at things our brains are natively
wired for,
like vision and speech recognition.
It would be a fairer comparison to compare the
arithmetic
capability of a workstation (~3 x 10^9/sec) to the
brain's
synapse firing rate. This implies that AI software
would
be designed to make efficient use of the CPU's
architecture,
or alternatively, a 'neural net card' was installed to
accelerate that type of calculation. This would be
similar
to graphics accelerators that are specialized for one
set of calculations.
A high end workstation, with four dual-core
processors, would
have a theoretical rate of 24x10^9/sec, or a factor of
4K below
the synapse rate equivalent.
When I make these type of calculations, I use a factor
of 30
below and above the raw synapse rate to allow for
uncertainty
in how much data one synapse firing represents.
Daniel
__________________________________________________
Do You Yahoo!?
Tired of spam? Yahoo! Mail has the best spam protection around
http://mail.yahoo.com
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:55 MDT