Integration quirks in GIF neuron models

[clinssen]

During a bug hunting session, me and @heplesser noticed that in one of the conductance-based GIF models, the spike-frequency adaptation and spike-triggered currents are updated outside of the inner GSL integration loop. In particular, they are updated before calling GSL; see

nest-simulator/models/gif_cond_exp.cpp

Line 493 in 07d2acc

for ( size_t i = 0; i < S_.stc_elems_.size(); i++ )

and the GSL call a few lines down from there.

A related issue in the current-based GIF model is that the exponential decay occurs first, and then the membrane potential is updated using the already-decayed value. See

nest-simulator/models/gif_psc_exp.cpp

Line 345 in 07d2acc

// exponential decaying stc and sfa elements

We should discuss whether this is indeed an issue, or whether it's just a minor implementation detail. Adding sfc and sfa within the inner loop could be done without too much trouble, but it is not clear where the updates (sfa += 1) should happen. As the neurons are not advertised as "precise", possibly these can be left in place.

In NESTML, this problem of mixing precise integration (using propagators) into the numerical integration routine is already addressed by nest/nestml#515.

[clinssen]

This issue is a duplicate of #1565.