The Effects of Biological Constraints on the Performance of Spiking Neural Networks

Abstract

Brain-inspired intelligence technology is always cutting-edge research in Artificial Intelligence (AI). These years, mimicking the properties of nerve impulses in the brain, a new type of deep learning network structure has been introduced-Spiking Neural Networks (SNNs). However, the properties of SNNs are still poorly understood, especially their potential biological plausibility. Here, we investigated Spiking Recurrent Neural Networks (SRNNs) obtained by parameters transformation. We investigated their performance and characteristics when achieving working memory tasks under biological constraints from the real brain. Finally, it was proved that the constraints introduced by us are biologically reasonable and can help to create SNNs with keeping both working memory capacity and biological plausibility.