Delay spiking neural networks
A delay spiking neural network (DSNN) may include a plurality of neurons arranged in a plurality of layers, with neurons spiking based on accumulation of delayed inputs. A first to spike neuron in an output layer may provide a result of the DSNN.
1. A method of determining delays for a delay spiking neural network (DSNN), the DSNN including neurons which determine output spikes based on accumulation over a period of time of delayed indications of received input spikes, with a delay associated with each received input spike, the method comprising:
iteratively applying training patterns to the DSNN, each application of the training patterns expected to generate a first spike by a particular neuron in an output layer; and
for each training pattern, in response to a first spike by a neuron in the output layer other than the particular neuron, backpropagating a positive error value for those neurons of the output layer other than the particular neuron and backpropagating a negative error value for the particular neuron, backpropagation of the positive error value resulting in non-negative increases in the delays and backpropagation of the negative error value resulting in non-positive increases in the delays.
2. The method of claim 1 , wherein the positive error value backpropagated for a neuron of those neurons of the output layer other than the particular neuron is based on a time difference between a time of a spike by the neuron prior to the first spike and a time of the first spike, and the negative error value backpropagated for the particular neuron is based on a sum of time differences between times of spikes prior to the first spike by those neurons of the output layer other than the particular neuron and the time of the first spike.
3. The method of claim 1 , further comprising:
storing indications of contributions of received input spikes in generating output spikes; and
using the stored indications in modifying the delays.
4. The method of claim 1 , wherein, for any training pattern, in response to a first spike by the particular neuron in the output layer expected to generate the first spike, ceasing iteratively applying training patterns to the DSNN.