Neuron simulation - who's doing it?

Rama Ratnam ratnam at leipzig.ks.uiuc.edu
Sat Nov 2 17:58:53 EST 1991


In article <1991Nov2.002242.9008 at cl.cam.ac.uk> amtw at cl.cam.ac.uk (Adrian  
Wrigley) writes:
> Sent to comp.ai.neural-nets but without any response:
> 
> > What is the current state-of-the-art in simulation of biologically
> > plausible neural networks?
> > Adrian Wrigley 
> 
> Who out there has a simulation of groups of neurons operating with pulses
> and local adaption algorithms?  I have been thinking of setting up a system
> using 1000+ neurons operating on speech spectra in real time, and was
> wondering how succesful simple (modified) Hebbian learning would be at
> recognising phonemes/words? (which is the most appropriate newsgroup for  
this?)
> --
> Adrian Wrigley
> Cambridge University Computer Laboratory, England.

I seriously doubt whether a simulator for biological networks
exists, that combines learning with the modelling. I may be
ignorant but i think this would be fairly deficult to develop.

I do have some info for you (if it is of any help)

1.There is a simulator from Caltech which i modified to suit
our research requirements. The simulator was originally
developed at Caltech by a group of people working on
the cat visual cortex. It runs on the Connection machine.

I  routinely simulate 24,000 neurons and approx 5 Million
synapses. This is pretty large and ties up the CM for quite 
a while  (about 1.5 hours per 8k CM sequencer).
(I'd believe that this is currently the largest scale simulation
going around)

2.Another simulator is BIOSIM developed by the SANS group at
the Royal Institute of Technology in Sweden. This is a really
detailed model based on Compartments and the Hodgkin-Huxley 
equations. It also runs on the CM, but a serial version called
SWIM (also developed by the group) is useful for small 
groups of neurons. Both simulators model channels.

3. GENESIS, also developed at Caltech is another simulator
that runs on serial machines. It has an X interface. I have never
tried it out, but a couple of my group members use it.

The caltech simulator (the CM one) works on populations and pathways,
does not simulate dendritic compartments, doesn't care too much
about channels and uses a leaky integrate-and-fire model.

Both have their plus points.
Simulating biological neural networks is a major headache.
The design issues are messy,  so i suggest that if you are interested,
lets correspond by email. I plan to write my own simulator soon (for the 
Connection machine).
My suggestion: If you are trying to simualte a few hundred neurons
with no more than a few thousand synapses, then you would be
stretching a serial machine pretty hard. Even a 1000 neurons is
quite a bit. AND if you want to add an adaptation module - phew!

-ratnam

--
==============================================================
Rama Ratnam,  Theoretical Biophysics, Beckman Institute
Dept of Biophysics, University of Illinois--Urbana-Champaign
ratnam at lisboa.ks.uiuc.edu
==============================================================



More information about the Neur-sci mailing list