Research On Neuromorphic System Based On Analog Circuit

With the rapid development of neural networks,it has shined in many fields,but at present,traditional neural networks mostly use software to run on computers.In order to achieve the same tasks such as image recognition,compared to biological brains,they consume a lot of hardware resources and energy.Based on this,the researchers proposed the concept of neuromorphic computing,which aims to use large-scale integrated circuits to simulate and implement the architecture of the neural system,to simulate biological neurons and neural systems.This article mainly carried out the following three aspects in the research of neuromorphic system.For the network algorithm model,a three-layer artificial neural network was trained for MNSIT handwritten numbers recognition,and the network's size and structure were changed to obtain different recognition rates.When the network scale is 784×784×10,the highest recognition rate is 98.43%.Then,the method of weight binarization and spiking frequency coding is used to transform artificaial neural network into a binarized spiking neural network with the highest recognition rate of 87.31%.In order to facilitate the subsequent neuromorphic hardware research,the memristor spiking computing scheme model is obtained by mapping the binary weighted network to the memristor cross array,and it is simulated using Matlab.The simulation results show that under the same network scale,the recognition rate of the memristor spiking computing scheme model is almost equal to that of the binary spiking neural network,The highest recognition rate is 87.28%,which means that the model works correctly.Based on the behavior simulation of the memristor spiking computing scheme model by using Matlab,using Cadence to simulationg the scheme.Building an integrated and fire(IF)neuron model,and modeling the memristor cross array,thereby constructing a two-layer neural network with a scale of 64×10,and using Cadence to simulate the network.The simulation results show that the neural network can run correctly and recognize input numbers.By using 10000 samples to test it,the final recognition rate is 52.86%,which is similar to the 52.43% recognition rate of the network algorithm simulation of the same scale.A 64×64×10 network is further built and tested with 1000 samples,and the recognition rate is 70.5%.The hardware feasibility of the algorithm is proved from the circuit simulation level.Based on the research of circuit simulation,the hardware design of the computational array of the neuromorphic system is carried out,and the hardware implementation of the memristor spiking computing scheme is explored.Using a 0.5?m CMOS integrated circuit process,a cross-array of memristors of different sizes is designed,which mainly includes two structures,1T1 R and 1D1 R.According to the current memristor technology,the process flow is divided into a front-end design and a back-end process design.The front-end design mainly includes the design of transistors,diodes,and cross-array select circuits,and the corresponding layout design has been successfully taped out and obtained chips.The subsequent process design mainly includes the integration of memristor.we completes the design of the memristor structure and the design and manufacture of the mask required for the subsequent growth process.These three research aspects from network training and transformation to neuromorphic system circuit simulation to part of hardware design,interlocking and progressive,prove the feasibility of the simulated neuromorphic system.