Energy-efficient time-series regression with spiking neural network

Abstract

We adopt a spiking neural network (SNN) for time-series regression in cerebral oximetry. The SNN outperformed other networks in the time-series regression when having a moving average filter. It is hypothesized that this is attributed to the noise shaping property of SNNs. With this in mind, we evaluate the performance of our SNN against other comparable networks: EfficientNet, long short-term memory, and fully connected network. The SNN has not only reduced network parameters, but also lower computational cost due to its inherent sparsity and the discrete nature of spiking neurons.