[Pcsim-users] Gamma distribution and negative weights

[Mathieu D. <mat...@li...>]

Hello PCSIMers,

I would like some advice on the best way to construct the reservoir of 
an LSM.

In the standard LSM literature, the weights of the intra-reservoir 
synapses are drawn from a Gamma distribution with parameters depending 
on the type of neurons connected (up to now I was using constant weights 
but I would like to see if using those weights increase performance).
I have found the BndGammaDistribution object in PCSIM and the relevant 
parameters in [1].

The problem is: how can I generate negative weights within the liquid 
(to connect an inhibitory neuron to another neuron)?
The Gamma distribution returns positive numbers so I have to do this by 
hand.

I have looked into the (matlab) circuit toolbox and found the code:
if synpar.W > 0
  synpar.W = sign(synpar.W) * bnd_gammarnd(abs(synpar.W), sh(1), 
abs(6*synpar.W), 1, m, 'exz A');
else
  synpar.W = sign(synpar.W) * bnd_gammarnd(abs(synpar.W), sh(1), 
abs(6*synpar.W), 1, m, 'inh A');
end

Under PCSIM, I use 4 SimObjectVariationFactory (one for each type of 
synapse: EE, EI, IE, II) and 4 ConnectionsProjection. Only the IE and II 
weights have to be negative. I was thinking at using the setFieldScale 
method in SimObjectPopulation. For instance for the II connections this 
would give:

liquid_syn[INH][INH] = 
ConnectionsProjection(self.liquid_sub_populs[INH], 
self.liquid_sub_populs[INH], syn_factory, 
PredicateBasedConnections(EuclidianDistanceConnectionPredicate(C[INH][INH], 
lambda_conn)))
tmp_pop = SimObjectPopulation( liquid_syn[INH][INH].idVector() );
tmp_pop.setFieldScale("W", -1);

However I'm not sure this is correct or if it is the best way.

Any advice?

Thanks in advance,
Mathieu

[1] W. Maass, T. Natschläger, and H. Markram. Real-time computing 
without stable states: A new framework for neural computation based on 
perturbations <http://www.igi.tugraz.at/Abstracts/MaassETAL:01a>. Neural 
Computation, 14(11):2531-2560, 2002. (PDF 
<http://www.igi.tugraz.at/maass/psfiles/130.pdf>, 886 KB). (hum, you 
probably know this one )