I recently found a bug in the SyncBuffer of Auryn (<0.7.x). This exclusively affects parallel simulations with more than one spike attribute. Specifically, these are simulations with more than one Connection class implementing short-term plasticity. I am still in the process of understanding the full implications of this problem, but there currently is a fix in Auryn 0.7-dev (
https://github.com/fzenke/auryn/commit/ ... aad250c5aa) which will be merged to the release version as soon as I am convinced the problem is fully contained. Until then any feedback on remaining issues is highly appreciated.
The bug seems to have a purely stochastic effect on synaptic conductance during steady state firing, but seems to introduce a systematic effect at transitions to high rates. The way this effect manifests is that it seems to make short-term depressing synapses slightly less depressing.
The code from doi: 10.1038/ncomms7922 (Orchestrated plasticity) is also affected by this. Here an illustration of the ramifications on the AMPA conductance of a cell before and after the SyncBuffer fix.
- conductance_difference.png (64.79 KiB) Viewed 13614 times
The message of the plasticity paper is unaffected by this. The effect can be easily counterbalanced by a small increase of exc input firing rates during stimulation or a small decrease in the time constant of synaptic depression (tau_d). The current develop branch of
https://github.com/fzenke/pub2015orchestrated readily implements one possible change. The input firing rate during stimulation is 50Hz instead of 35Hz (other, lower values might also work, but I did not try this) to compensate for the effect.
