Construction Of Binary Spiking Neural P System Models

Membrane computing(P-system)is a distributed computing model that imitates the structure and function of cells.Non-linear Spiking Neural P(NSNP)system provides new idea of constructing non-linear supervised learning networks,which is a kind of non-linear P-system.Due to high computing cost,supervised learning networks work badly in low power-consumption scenarios,such as mobile devices and hardware.Therefore,network compression is necessary,and Binary Neural Networks(BNNs)are proven as an efficient compression technology.But there are big accuracy gaps between BNNs and Convolutional Neural Networks(CNNs).The main work and contribution of this paper are as follows:1)According to the NSNP system,we derive SNP-like neurons and construct Convolu-tional Spiking Neural P(Conv SNP)system based on SNP-like neurons.Comparative experiments on common image datasets prove that Conv SNP shows higher performance and stability on small and medium datasets than traditional CNNs.2)Based on BNNs,we construct Binary Convolutional Spiking Neural P(Bin Conv SNP)system.Comparative experimental results on common image datasets show that Bin Conv SNP has a higher recognition accuracy compared with BNNs.It is proved by experiments that the model size of Bin Conv SNP is 1/30 of CNNs’,and the calculation speed increased by 57 times.3)We analyze the flaws of Batch Normalization layer in BNNs,and propose Two-peak Batch Normalization(TPBN)layer,then improve Bin Conv SNP by using TPBN layer.It is confirmed by experiments on common large-scale image data sets that Bin Conv SNP with TPBN layer can increase by 5-7% in accuracy compared with BNNs.Our study combined CNNs with NSNP system successfully to construct a new type of nonlinear supervised learning network.And we construct a non-linear supervised learning network model which can work efficiently in resource-limited scenarios by compressing the new network with binary network technology.The proposed Bin Conv SNP system in this study can be quickly computed in low power-consumption scenarios,so it is more suitable for edge computing than the full-precision CNNs.