Asynchronous Spiking Neural Network Design And Implementation Based On MESH Structure

With the 50 years of continuous development of Moore's Law,the integrated circuit industry has gradually developed in a large-scale direction.The resulting area and power consumption problems have become increasingly prominent.For technical and economic reasons,it is widely believed that it will end around 2020.With the gradual end of Moore's Law,how will the semiconductor industry develop in the post-Moore's Law era? Different industries have different views on this issue.We are focusing on brain-like computing.In this paper,a construction method of asynchronous spiking neural network based on asynchronous MESH grid structure is proposed,which adopts asynchronous micro-pipeline control mechanism and multi-core parallel computing design.Several neuron models and pulse sequence coding methods of spiking neural network are discussed.The current-based LIF model and time-coded pulse coding method are selected.The neuron model is analyzed and the neuron model algorithm is explained.For the application scenarios of digital image recognition,we use the Tempotron algorithm as the basis for implementing neuron units.The neuron model algorithm is optimized and implemented for hardware circuits.Then,based on this,the digital image recognition of 0~9 is realized by the upper computer.The experimental results were collected and verified.Compared with the traditional neural network implemented by software,it is more in line with the human brain drive,event-driven way,faster and lower power consumption.Finally,we apply the implementation mechanism of asynchronous spiking neural network to MESH network,and realize and simulate the design of asynchronous spiking neural network based on MESH structure through Xilinx Vivado development tool.