Research On Optical Spiking Neural Network Based On Reconfigurable Interconnection

The rapid development of the information age has led to great changes in computing systems.High-complexity artificial intelligence systems constantly put forward higher requirements for computing speed and energy efficiency,making traditional computing systems based on von Neumann architecture gradually encounter the bottleneck of performance improvement.In order to solve this problem,researchers introduce information processing mechanism based on optical spiking,build a new computing architecture,and establish optical spiking neural networks.With the advantages of photon's high communication bandwidth,high switching speed and low crosstalk,optical spiking neural networks can achieve ultra-fast spiking processing with high energy efficiency and high interconnection density.This thesis describes the evolution and development prospect of optical spiking neural network,mainly studies the reconfigurable interconnection in this field.The thesis compares three mainstream reconfigurable interconnection schemes based on microring resonator,Mach Zehnder interferometer and phase change material,and on this basis,mainly exploring the interconnection scheme based on microring resonator.Aiming at the problem of weight loss caused by inter-ring coherence effect in this scheme,the research includes two aspects.1)Research on optimizing the channel density of microring weight banks.The channel density determines the interconnection density of neural networks and greatly affects the network performance.Existing research has explored the influence of inter-ring coherence effect on weight power penalty,and analyzed the channel density of microring weight banks.However,the inter-ring coherence effect is not the only influencing factor of weight power penalty,and an analysis based on a single factor can not be able to accurately evaluate the channel density of microring weight banks.In this thesis,other physical parameters of microring weight banks are taken into consideration,and the influence of physical parameters on weight power penalty is explored.Furthermore,an optimization scheme for the channel density of microring weight banks is proposed.The proposed scheme can achieve higher channel density with the same weighting performance.Simulation results show that by appropriately configuring the physical parameters of microring weight banks,the channel density can be increased by 34.5%,which is about 2.5times of the existing scheme.2)Research on optimizing the accuracy of optical spiking neural networks based on microring weight banks.The microring weight bank has a unique inter-ring coherence effect,which makes the actual weight range of the bank lower than the ideal range,resulting in the loss of accuracy when the network is loaded into the hardware system.This thesis proposes to predict the actual weight range of microring weight banks by modeling and simulation,and take the calculated weight result as the constraint of network training.This method can ensure that the weight of the trained network is within the actual weight range of microring weight banks,so as to improve the accuracy of optical spiking neural networks.The test results on MNIST dataset show that the proposed method can improve the recognition accuracy by up to 1.6% compared with the conventional training method without considering the characteristics of microring weight banks.