Dear Andreas,

>> In particular, the parameters drvifall (affects the time
>> constant and synaptic weight; this parameter first, because it is
>> more sensitive) and drviout (affects mostly the synaptic weight)
>> are adjusted until a target rate of the postsynaptic neurons is
>> reached. As these parameters can be configured only row-wise for
>> synapses, firing rates are averaged across all hardware neurons
>> during calibration. Due to this protocol, synapses are
>> calibrated to generate EPSPs with comparable impact on the firing
>> rates of postsynaptic neurons (i.e. area under EPSP), but not
>> necessarily with comparable time constants.
> Ok, so as far as I understand it the LIF parameters "w" (synaptic
> weight) and "tauM" (membrane time constant with tauM = cM / gL)
> influence the hidden "tauE" (excitatory synapse time constant)
> parameter. That's also what some quick experiments on the hardware
> showed.
>
> So in case "tauM" and "w" stay fixed I can calculate "tauE" and
> adapt the other parameters (e.g. threshold and leak potential
> "eL"/"eTh") without implicitly changing "tauE". That's enough for me.

Note that the parameters drvifall and drviout are additional  
parameters to w and tauM.
Drviout mostly affects the height of EPSPs, like w, but drvifall  
affects both tauE and the height of EPSPs.
I have not measured any dependencies of w and tauM on tauE, yet, but  
these dependencies may be artifacts, because the neuron on hardware is  
not exactly an LIF neuron (see also [1]).
Further investigations would be interesting.

>> As it is difficult to extract the synaptic time constant from the
>> membrane potential, we can not provide you with numbers on how
>> correlated time constants and area under EPSPs are.
> Well, that turns out to be really simple: The estimation of tauE
> from a membrane potential trace is a matter of few seconds. I've
> simply hooked my (rather naive) Downhill-Simplex implementation onto
> my high-speed single-neuron AdEx integrator and defined the RMSE
> between simulated and reference membrane potential trace as cost
> function. See attachment for the results.

Sorry, I mixed up the measurement and calibration of tauE.
For example, [1 and 2] already measured tauE and calibrated this time  
constant for single neurons, but so far we did not find an appropriate  
method to calibrate tauE for all synapses on the chip.
As our experiments were never sensitive to tauE, we have been happy  
with the firing rate calibration of drviout and drvifall.

Regards,

Thomas

[1] Petkov Diploma thesis:  
http://www.kip.uni-heidelberg.de/Veroeffentlichungen/details.php?id=2635
[2] Brüderle PhD thesis:  
http://www.kip.uni-heidelberg.de/Veroeffentlichungen/details.php?id=1917

########################################################################

To unsubscribe from the KIP-SPIKEY-USERS list, click the following link:
https://listserv.uni-heidelberg.de/cgi-bin/wa?SUBED1=KIP-SPIKEY-USERS