Jethro's Braindump

Synaptic Current Model

Synaptic currents are generated by synaptic currents triggered by arrival of presynaptic spikes Sj(l)(t). Spike trains Sj(l)(t) are denoted as a sum of Dirac delta functions Sj(l)(t)=sCj(l)δ(ts), where s runs over the firing times Cj(l) of neuron j in layer l.

A good first-order approximation of the synaptic current is one of exponential decay. Synaptic currents are also assumed to sum linearly.

dIi(l)dt=Ii(l)(t)τsynexp. decay +jWij(l)Sj(l1)(t)feed-forward +jVij(l)Sj(l)(t)recurrent 

A single LIF neuron can be simulated with 2 linear differential equations whose initial conditions change instantaneously when a spike occurs. Combining the reset term with the equation for the Leaky Integrate-And-Fire model, we get:

dUi(l)dt=1τmem((Ui(l)Urest)+RIi(l))+Si(l)(t)(Urestϑ)

The solutions to Equations eq:scm and eq:lif_with_reset are approximated numerically by discretizing time, and expressing the output spike-train Si(l)(n) of neuron i in layer l at time-step n as a non-linear function of the membrane voltage Si(l)(n)Θ(Ui(l)(n)θ) where θ is the Heaviside step function, and θ is the firing threshold.

Setting Urest=0, R=1, θ=1, and using some small simulation time step δt>0, we get:

Ii(l)[n+1]=αIi(l)[n]+jWij(l)Sj(l)[n]+jVij(l)Sj(l)[n]

with decay strength αexp(Δtτsyn). Equation eq:lif_with_reset can then be expressed as:

Ui(l)[n+1]=βUi(l)[n]+Ii(l)[n]Si(l)[n]

with βexp(Δtτmem). These two equations characterise the dynamics of a RNN. Specifically, the state of neuron i is given by the instantaneous synaptic currents Ii and the membrane voltage Ui.

References

Bibliography

Neftci, Emre O., Hesham Mostafa, and Friedemann Zenke. n.d. “Surrogate Gradient Learning in Spiking Neural Networks.” http://arxiv.org/abs/1901.09948v2.