Refractoriness always in threshold.py

[mstimberg]

This implements the automatic addition of "not_refractory" to the threshold condition. It also shows a problem with this approach, though: PoissonGroup doesn't use the refractory period but since not_refractory is reset in the state updater, it needs a crazy almost empty state updater to reset the not_refractory variable at every time step. Obviously we want to handle this better, maybe specific templates for thresholds/state updates without refractoriness.

[coveralls]

Coverage remained the same when pulling 54de343 on refractoriness_always_in_threshold.py into 00e7ecd on master.

[thesamovar]

I'm wondering if we can make PoissonGroup not a group, but say PoissonSource derived directly from SpikeSource and not Group.

[mstimberg]

I'm wondering if we can make PoissonGroup not a group, but say PoissonSource derived directly from SpikeSource and not Group.

I think there was a good reason for it, but I don't quite remember what the reason was :) One advantage of it being a group is that you can access the rates variable as a state variable. With this comes the possibility of using string expressions in synapses referring it, which might be quite neat. Something like: S.w = 'w0 + 1./rates_pre*w1'. It also means you can set rates with a string expression which allows its use in standalone devices.

[thesamovar]

Could we do this with a Group instead of a NeuronGroup? And if so, it shouldn't have any refractoriness stuff in there?

[mstimberg]

Could we do this with a Group instead of a NeuronGroup? And if so, it shouldn't have any refractoriness stuff in there?

It's already a Group, not a NeuronGroup but it uses a Threshold object with threshold 'rand() < rates * dt' and because the threshold template also sets not_refractory, we do need all the refractoriness stuff. I think we can either make the threshold/state update mechanism more general/flexible (this would also allow to remove refractoriness for models that don't use it) or simply use a specific template for PoissonGroup.

[thesamovar]

Ah, I see. I think I'd probably prefer to make refractoriness optional in models but only if it doesn't make everything rather complicated.

[mstimberg]

Ah, I see. I think I'd probably prefer to make refractoriness optional in models but only if it doesn't make everything rather complicated.

Maybe we can do it on the template level, i.e. via a keyword provided to the template which then conditionally includes the refractoriness bit?

[thesamovar]

Sounds good! Can we detect it automatically actually? i.e. if no user code touches the refractory variables then don't bother to have them.

[mstimberg]

Can we detect it automatically actually? i.e. if no user code touches the refractory variables then don't bother to have them.

I think we can quite simply: If the user does not specify the refractory keyword in the NeuronGroup constructor, we don't have refractoriness!

[thesamovar]

OK should we add an issue for this then?

[mstimberg]

Issue is here: #125