A Hardware Accelerator of the Convolutional Spike Neural Network Based on STDP Online Learning

Convolutional Neural Networks (CNNs) excel in object recognition but are computationally intensive and unsuitable for online learning. Spiking Neural Networks (SNNs) [1] have low computational requirements, fast processing, and low power consumption, making them suitable for online learning. This paper proposes a hardware accelerator for a Convolutional Spiking Neural Networks (Conv-SNNs) which combines the strengths of CNNs and SNNs by introducing spiking mechanisms into the convolutional layers. It utilizes Spike Timing Dependent Plasticity (STDP) and Reward-STDP (R-STDP) algorithms for training, achieving 95% accuracy on MNIST dataset. In addition, it also supports the few-shot learning. Implemented on Xilinx's ZCU102 FPGA at a 100MHz clock frequency, the system achieves 0.16s for inference and 0.177s for training per image, 16x faster than CPU. The power consumption is 3.979W, two orders of magnitude lower than CPU.