questions about networks

[rohrkemper]

I will outline what I propose to do. It would be helpful if you could tell me how well suited NEURON is for this particular type of problem and how I can get started.

I would like to have a network of several hundred or possibly more conductance-based leaky integrate and fire neurons. I'm also interested in using parallel NEURON to extend the number of neurons.

First, I have a connectivity matrix that is generated with some MATLAB routines. I would like to build this network in NEURON and to record the spiketrains in a format that is again readable by MATLAB for correlation analysis. I would like to be able to easily change the type of neuron model that is used. Is a conductance based neuron model available? I noticed that regular LIF is mentioned.

Regards,
Bobby

[ted]

It would be helpful if you could tell me how well suited NEURON is for this particular type of problem

Eminently.

As part of a multi-authored paper on methods for modeling networks of spiking neurons
(Brette et al., submitted for review), we used NEURON to replicate the model
described in
Vogels TP, Abbott LF. (2005) Signal propagation and logic gating in networks of
integrate-and-fire neurons. J. Neurosci. 25: 10786-10795.
The model net consisted of 4000 model cells (3200 excitatory, 800 inhibitory), connected
randomly (P(a projects to b) = 0.02). We ran benchmarks using four different
implementations. The implementations differed only in the way that cellular properties
and synaptic effects were represented.
1. leaky integrate and fire cells with "current-based" synapses ("CUBA" model)
2. leaky integrate and fire cells with "conductance-based" synapses ("COBA" model)
3. Hodgkin-Huxley cells (modified from Traub, R.D. andMiles, R. (1991) Neuronal
Networks of the Hippocampus. Cambridge University Press, Cambridge, UK) with
"conductance-based" synapses
4. integrate and fire cells with "voltage-jump" synapses

The code was factored into "network-specific" and "cell- and synapse-specific"
components, which makes it easy to change network architecture or the nature of the
component cells and synapses independently of each other. It runs without modification
on standalone single processor PCs or Macs under OS X, MSWin, or UNIX/Linux,
as well as on parallel hardware.

The benchmarks, which included tests on a Beowulf cluster and the Cray XT3 at the
Pittsburgh Supercomputing Center, showed linear speedup with the number of
processors "as long as each processor has enough to keep it busy." In other
words, nets nets of cells with conductance-based synapses benefitted a lot from
parallel hardware (up to 800 CPUs on the Cray), but nets of the simplest integrate and
fire cells with "current-based" synapses executed very quickly even on standalone
single processor hardware and so gained little by parallelization.

how I can get started.

Start by learning the basics of modeling with NEURON, and pay special attention to
chapters 9 and 10 of The NEURON Book. Then get the source code for the models
described above and modify it, or borrow from it, as needed to set up your own net.
The code has been deposited in ModelDB as a "private entry," and will be made
"public" after the manuscript has been accepted for publication.