Neuron reset after spike event and synaptic update

Hi everyone,

General description

I’m playing a with minimal triple-STDP learning rule.
For each postsynaptic and presynaptic neuron, there are low-pass filters of spikes, which contribute to the change of synaptic plasticity.

\begin{array}{rl} r_1' &= -r_1/\tau_{+} + S_{pre} \\ o_1' &= -o_1/\tau_{1-} + S_{post} \\ o_2' &= -o_2/\tau_{2+} + S_{post} \\ w' &= r_1(t) o_2(t-\epsilon)A^{+}S_{post}-o_1(t)A^{-}S_{pre} \end{array}

where S_{pre} and S_{post} are presynaptic and postsynaptic firing rates.
The small constant \epsilon is added to indicate that synaptic weight w is updated before update o_2.

1. Looks like the most efficient implementation

The most efficient way is to implement r_1, o_1, and o_2 equations on the neuron side:

equ ="""
dv/dt = ....
dr1/dt = -r1/tau1p : 1
do1/dt = -o1/tau1m : 1
do2/dt = -o2/tau2p : 1
npop = NeuronGroup(... equ, reset='v=vres; r1+=1; o1+=1; o2+=1' ....)

and then use these variables in synaptic equations.

s = Synapses( 'w:1', 
    on_pre='w = w-o1_post*Am',
    on_post='w = w+r1_pre*o2_post*Ap', ....)

But it wasn’t clear to me whether the neuron reset will be done after the synaptic update or not.

2.Very slow implementation but with the persistent result

I fixed this by moving equations for r_1, o_1, and o_2, on the synaptic side and updating all variables after the synaptic weight update.

s = Synapses( """
    w : 1 
    dr1/dt = -r1/tau1p : 1 (event-driven)
    do1/dt = -o1/tau1m : 1 (event-driven)
    do2/dt = -o2/tau2p : 1 (event-driven)
         w = w-o1_post*Am
         r1+= 1""",
         w = w+r1_pre*o2_post*Ap
         o2+=1""", ....)

However, this implementation is VERY slow because instead of a few thousand differential equations, I got a few million.

Finally question

Is there any way to ask Brian to do

  1. Update all neurons
  2. Trigger all events
  3. update all synapses
  4. reset all neurons.

The default schedule is ['start', 'groups', 'thresholds', 'synapses', 'resets', 'end'], where 'groups' refers to the update of variables defined by differential equations, and 'synapses' refers to the on_pre and on_post statements. So e.g. on_pre will be executed before the reset. You can verify this with scheduling_summary. If you need to change this, you can either change the global schedule, or move individual objects around. For example, s.pre.when = 'after_resets' would move the on_pre execution to the slot following the reset.

I’m not quite sure I understand – you mean without running a simulation? In principle, you could trigger the individual code_objects responsible for doing the tasks, but I’m not sure whether that’s what you have in mind?

Thank you, @mstimberg. It’s much clearer now.

1 Like