From: Rüdiger Koch (rkoch@rkoch.org)
Date: Sat Apr 06 2002 - 09:51:46 MST
Very good paper. I don't find it supports Eugene's estimate of 3TeraIPS for
a single neuron with 2x20 synapses, however. Quite contrary. It
essentially says that you can view a neurons as a leaky integrator of
incoming spikes. Some parameters of that integration are the concentration of
certain ions (Na+, Ca+, K+). According to that paper, synapse strength and
spike timing are the dominant quantities for spiking. Spike timing is the
dominant influence in Hebbian learning.
This is *exactly* what we're doing with Amygdala. We can now run around 1,000
neurons/CPU with 1,000 dendritic synapses each in real time. If you think
that the Koch/Segev paper describes neurons detailed enough and then 1 neuron
eats roughly 0-5 MIPS, depending on the amount of inputs it gets.
The paper doesn't really cover Nematode neurons which output a steady voltage
which depends on it's inputs. Mammal neurons as described in the paper are
considerably more complex because spike timing is the most important measure
both for processing and for learning.
On Saturday 06 April 2002 17:12, Eliezer S. Yudkowsky wrote:
> May I recommend:
> "The Role of Single Neurons in Information Processing"
> http://icnc.huji.ac.il/Files/Koch_Segev_NN2000.pdf
> for those wondering why uploading is so computationally expensive.
>
> -- -- -- -- --
> Eliezer S. Yudkowsky http://singinst.org/
> Research Fellow, Singularity Institute for Artificial Intelligence
-- Rüdiger Koch http://rkoch.org Mobile: +49-179-1101561
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 09:13:16 MST