4. Learning with spikes¶
To be able to learn in spiking neural networks (SNN) one needs to update the weights between neurons, as in all types of neural networks. As you probably know, you need smooth, differentiable functions to apply popular algorithms like gradient descent and backpropagation.
However, spiking neurons do not have smooth activation functions (a spike either happens or doesn’t). This page aims to explain - in an informan manner - how we can train SNNs in Norse irregardless of their non-differentiable nature. The general approach is described in better detail by Emre O. Neftci, Hesham Mostafa, and Friedemann Zenke.
Before you read further though, make sure you are familiar with how PyTorch works with Autograd and backpropagation. Norse is basically applying the same principle.
4.1. Training Spiking Neural Networks¶
Many solutions have been attempted to solve this problem with varying success. Here we will only cover the surrogate gradient approach, and illustrate how it is implement in norse.
Our approach builds on the SuperSpike method proposed by Steven K. Esser et al. (Convolutional networks for fast, energy-efficient neuromorphic computing) and further elaborated in Friedemann Zenke and Surya Ganguli.
Neurons work in the way that they update their membrane equations with incoming currents from pre-synaptic neurons. If the incoming currents exceed a threshold, the post-synaptic neuron releases a spike. Hhis can easily be expressed in code:
if membrane > threshold: spike!
However, what happens when you take the gradient of that? It will be
zero for the most part because
membrane < threshold, meaning that
the neuron does not influence the output at all.
But sometimes the current goes above the threshold, you get an
activation, and the gradient changes!
To account for that, we can “pretend” that the gradient doesn’t
have this awkward sudden shift.
Instead, we can look at the numerical state of the neuron and then
use that as an indicator for how much the neuron influenced the
Given the neuron membrane potential (\(U\)) and the neuron firing threshold (\(v\)), then this is a simplified version of the SuperSpike surrogate partial derivative for some activation function (\(\sigma\)):
In the SuperSpike algorithm, we look at the difference between the neuron membrane and the firing threshold. If, say, the neuron membrane voltage is much higher than the firing threshold, we know that the neuron will fire. But too far away from that threshold indicates that the contribution of the neuron is unimportant because it would require a large modification to that particular neuron to not impact the output.
Conversely, if the the neuron membrane voltage is much lower than the threshold, the neuron is probably not going to fire, and the gradient contribution is equally low.
And that’s it! SuperSpike permits the calculation of gradients for non-differentiable functions. Which, in turn, permits us to use the native autograd properties of PyTorch.
An implementation of the SuperSpike algorithm in Norse can be found in the threshold.py module.