From: Eugen Leitl (eugen@leitl.org)
Date: Sat Apr 06 2002 - 07:30:39 MST
On Sat, 6 Apr 2002, Rüdiger Koch wrote:
> As far as I am concerned uploading C.Elegans was solved by Ferrée
> ( http://www.csi.uoregon.edu/members/ferree/publications.html ) in 1996 until
> someone explains what's wrong with his solution. The fruit fly is orders of
I presume you refer to the following four publications from 1997 to 1999:
T. C. Ferree and S. R. Lockery (1999). Computational rules for chemotaxis
in the nematode C. elegans.
Journal of Computational Neuroscience6: 263-277.
T. C. Ferree and S. R. Lockery (1998). Chemotaxis control by linear
recurrent networks.
Computational Neuroscience: Trends in Research: 373-377. Plenum
Press.
T. M. Morse, T. C. Ferree and S. R. Lockery (1998). Robust spatial
navigation in a robot inspired by C. elegans.
Adaptive Behavior: Special Issue on Biomimetic Navigation Models6:
391-408. MIT Press.
T. C. Ferree, B. A. Marcotte and S. R. Lockery (1997). Neural network
models of chemotaxis
in the nematode C. elegans. Advances in Neural Information Processing
Systems 9: 55-61. MIT Press.
While certainly interesting papers this doesn't qualify because he's using
a simplified model (sans direct input from neuroanatomy, nor cell models
including input from the genome) explaining one aspect of behaviour. It
doesn't account for the entire behaviour repertoire, it doesn't show
individual variations (as verified by laser ablation in vivo and in
machina plus mutants), in other words is as much a worm upload as a 18th
century android is a human. (I'm overstating my case for sake of an
illustration).
> magnitude more difficult (200,000 spiking neurons vs 302 graded signal
> neurons). IMO it should be possible to emulate a fruit fly on a PC cluster of
> 20-100 processors.
I'm a big boggled by these numbers, I must admit. You got bitten by Ray
Kurzweil, or something? I would put the 10^2..10^3 CPU milestone for the
biologically realistic simulation of a nematode, and at an emulation speed
of 1:10^3..10^4, i.e. thinkably far removed from realtime. Assuming you
knew what you have to emulate, which we don't, currently. And while above
resources would be more than sufficient for a parametrized simulation, you
first need to have a fully detailed simulation before you can start to
abstract.
While we have the worm mapped down quite precisely (though no contiguous
data sets of an adult individuum exist), this is not true for any higher
organisms, which are also far less deterministically wired, and have a
much more complex behaviour set and live in far more complex environments
than the lowly worm grazing off a petri dish's bacterial lawn (which is
not exactly its native environment, but one has to start somewhere).
> ASCI White is most likely sufficient to emulate a Fruit Fly. Blue Gene (due
> 2004) should be sufficient to emulate a mouse. They are general purpose
You're confusing computational equivalent and biologically realistic. A
100 kNode Blue Gene filled with the right code and data might be capable
of mighty deeds (we don't know exactly how mighty), but this doesn't
happen as long as we don't have that magic conglomerate of code and data,
which is necessarily a great deal more abstract than just digitized
neuroanatomy. We don't have methods to generate such abstractions yet.
Which is the whole point of starting low, and generating methods. And the
worm is an pretty amitious first target already.
> computers. If the same effort would be put into a neuro machine then the
> hardware side of emulating a human brain is likely in reach now. I am saying
> most likely because we still don't know if neurons are way more complex as we
> think they are (e.g. Penrose).
Penrose is smoking heavy crack here, but biological neurons are not the
caricatures you deem them to be.
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 09:13:16 MST