Connectionists: Gradient Descent for Spiking Neural Networks

[Terry Sejnowski]

Spiking network models are not differentiable because of discontinuities
at the spike times.  This has prevented learning in spiking networks by
gradient descent.

We have a way to overcome this problem and show that it is possible
to train spiking recurrent networks to solve temporal sequence problems 
over a wide range of time scales using backpropagation through time:

https://arxiv.org/abs/1706.04698

This can be applied to any spiking neuron model including the Hodgkin-Huxley 
model, leaky integrate and fire neurons, the Morris-Lecar model and the 
FitzHugh-Nagumo model.

Terry

-----