The spiking neural network(SNN)is the third generation of neural networks,valued for its biological interpretability,hardware compatibility,and low power consumption.While field programmable gate array(FPGA),based SNN...The spiking neural network(SNN)is the third generation of neural networks,valued for its biological interpretability,hardware compatibility,and low power consumption.While field programmable gate array(FPGA),based SNN accelerators show promise with their intelligent design and low power consumption,existing models lack a crucial spiking encoding module.This leads to increased pre-processing time on the CPU.To address these issues,this paper proposes an FPGA-based SNN accelerator with a spiking encoding module.This innovation reduces CPU pre-processing time by threefold without adding FPGA processing time.Achieving 92.62%accuracy on the MNIST dataset using LIF spiking neurons at 0.675 W,the accelerator outperforms CPU and GPU,being 14×faster and 2.5×faster,respectively,and more power-efficient with minimal resource usage.Given its pixel input,it can be seamlessly embedded in real-time classification systems for efficient classification and detection.展开更多
基金supported in part by the National Natural Science Foundation of China(Nos.62104176 and 62171347)in part by the Fundamental Research Funds for the Central Universities(No.XJSJ23083)+2 种基金in part by the Proof of Concept Foundation of Xidian University Hangzhou Institute of Technology(No.GNYZ2023XJ0409-1)in part by the Shaanxi Higher Education Teaching Reform Research Project(No.23ZZ014)in part by the China Postdoctoral Science Foundation(No.2019M663637).
文摘The spiking neural network(SNN)is the third generation of neural networks,valued for its biological interpretability,hardware compatibility,and low power consumption.While field programmable gate array(FPGA),based SNN accelerators show promise with their intelligent design and low power consumption,existing models lack a crucial spiking encoding module.This leads to increased pre-processing time on the CPU.To address these issues,this paper proposes an FPGA-based SNN accelerator with a spiking encoding module.This innovation reduces CPU pre-processing time by threefold without adding FPGA processing time.Achieving 92.62%accuracy on the MNIST dataset using LIF spiking neurons at 0.675 W,the accelerator outperforms CPU and GPU,being 14×faster and 2.5×faster,respectively,and more power-efficient with minimal resource usage.Given its pixel input,it can be seamlessly embedded in real-time classification systems for efficient classification and detection.
