Spiking Neural Network with plasticity in the time domain recovers temporal information from a noisy pattern using reference spikes

Abstract

Neurons in the brain communicate with each other by sending trains of spikes that can encode information using the timings of the spikes. Spiking Neural Networks (SNNs) are biologically plausible neural networks that can model this transfer of information and that, by incorporating plasticity, can detect repeating patterns that may be embedded in a train of spikes with a noisy background. Although existing model networks are capable of learning to detect the occurrence of repeated patterns, they are not tailored to recover the entirety of the time sequence of the spikes in the pattern. Here we present a network that, in addition to typical parameters of plasticity such as weights and time delays, uses a new plasticity on the time domain that can recover most of the spikes in a pattern. This new plasticity acts by modifying the timings of reference spikes that come from a hypothesized upstream network that could potentially encode preexisting memories. The model is shown to be robust under several noise perturbation scenarios, and its overall performance demonstrates the benefits of using reference spikes to improve the temporal information processing ability of SNNs.