Can GeNN realize dual-exp synapse with "create_weight_update_model" method?

[OOOOOmoi]

I'm trying to implement dual-exp synapses via create_weight_update_model(), where the synapse_dynamics_code word defines the update of the synaptic weights at each time step, like this:

But unlike what I expected, it seems that the stimulus was not delivered to the postsynaptic neuron:

There area two LIF neuron and one synapse in the demo and the synapse model is DeltaCurr. The top of the figure shows the membrane potential changes in the presynaptic neuron, the middle shows the membrane potential changes in the postsynaptic neuron, and the bottom shows the recorded synaptic current.

[OOOOOmoi]

OK, I test the demo on another device and get the result below:

This dual-exp synapse worked but not every spike seemed to work, like the ones marked by the red box in the figure, which is strange

[OOOOOmoi]

I think I seem to have some understanding of what the problem is, the problem is in the specific value judgment of t-st_pre, after I changed the judgment condition to <= 0.9 * DT the synapse was able to pass all the spikes:

But if I change the condition to <= 0.0 then there is still a portion of the spike that cannot be passed:

[tnowotny]

This looks like a typical floating-point comparison issue. Due to potential rounding issues, it is unsafe to compare floating point variables with ==. The safe way is to compare abs(t-st_pre) < 0.1*DT or something similar. Note, the rounding errors can go both ways, hence the use of abs.

Another thought, it would be more elegant (and maybe slightly faster) if the increment of x was applied within the pre_spike_syn_code snippet (using GeNN 5.x naming here):

pre_spike_syn_code="x += 1.0;"

If done this way, there is no need for comparing t and st_pre.

[OOOOOmoi]

OK, thanks

[tnowotny]

As @neworderofjamie just pointed out to me the most efficient way is to use the dynamics of the postsynaptic model instead of synapse dynamics. This can be done as you are solving the equations
$\dot{x} = - \frac{x}{\tau_r}$
$\dot{g} = -\frac{g}{\tau_d} + x$
which are linear in terms of the input variable $x$, so can be swapped with the summation of incoming inputs $x_i$ from neurons $i$.

Hence, you can have a simple synapse with pre_spike_syn_code="addToPost(weight);" and no synapse dynamics.
Then, combine this with a postsynaptic model roughly like so:

postsyn_dual_exp = pygenn.create_postsynaptic_model(
"dual_exp",
params=[("taur", "scalar"), ("taud", "scalar")],
vars=[("g", "scalar")],
sim_code= """
    injectCurrent(g);
    g += (-g/taud + inSyn)*dt;
    inSyn += -inSyn/taur*dt; 
"""
)

[tnowotny]

The advantage is that now, the per-timestep calculations happen in the postsynaptic model, which has number of postsynaptic neurons instances, instead of in the weightupdate model, which has (number pre)*(number post) instances.

[OOOOOmoi]

Yes, this solution is more elegant that, thanks.