Compressed Sensing Spiking Neural Network for Sparse Signal Classification

The spiking neuromorphic processors designed specifically for brain-inspired spiking neural networks (SNNs) are in increasing demand with the advantages of energy efficiency. Due to the limitation of on-chip resources of edge devices, large-scale input signals are typically downscaled before being processed on SNN processors. In this work, we propose the compressed sensing spiking neural network (CSSNN) method for sparse signals, which combines the compressed sensing (CS) process with the SNN classifier. This method achieves an end-to-end structure for data compression and SNN classification. As a result, the overall computation complexity is largely compressed with minimal classification accuracy loss. The input data size of the MNIST dataset is compressed to 256 and that of the N MNIST and DVS Gesture datasets is 512. The total number of operations (OPs) of the network decreases by 65.24%,72.52% and 81.11%. The model size is compressed by 63.35%, 62.02%, and 62.75% respectively, without accuracy loss compared with the baseline.