Embedded Brain-Like Perception And Computing Based On Event-Driven Method

In recent years,neural networks have achieved rapid development and outperformed the original handcrafted feature design methods in many domains.However,the traditional artificial neural network(ANN)has a large amount of computation,so it is not suitable to be deployed on the embedded edge computing platform with low cost and limited resources.As an important implementation of brain-inspired algorithms,spiking neural network(SNN)supports the processing of time-domain information,and the sparseness in the network can reduce the amount of computation and power consumption.Neuromorphic chips can deploy spiking neural networks with power consumption in the milliwatt range,but they are expensive and inflexible.Software simulation can be deployed on existing computing platforms with greater flexibility.However,the traditional software simulation adopts the clock-driven method,which has the problems of many invalid calculations,limited simulation accuracy and failure of lateral suppression.In response to these problems,this topic is devoted to the research of event-driven spiking neural network simulation framework and the application of brain-like computing on embedded platforms.The main research contents are as follows:(1)Aiming at the high complexity of a single update of the event-driven simulation framework,two innovations of population computing and pre-filtering are proposed.Since the update interval of neurons within the population is consistent,the population computing can reuse the decay factors to avoid repeated calculations.Pre-filtering is to use the judgment condition of low cost for screening.If the condition is not met,it means that the neuron will not emit pulses,so there is no need for subsequent calculations.In addition,engineering optimizations such as function inlining,rewriting priority queues,and lookup table acceleration are used to further improve computing efficiency.The reconstructed framework is named Evt SNN.(2)The concept of steady-state potential is proposed and the relevant formula is derived,and the weight factor is adjusted according to the formula,which effectively solves the problem of sudden changes in the number of pulses in different experimental scenarios.Through performance analysis experiments,ablation experiments,benchmark experiments and unsupervised learning tasks,the simulation speed and accuracy of multiple frameworks are compared,and the acceleration effect of each optimization item in the Evt SNN framework is evaluated.In benchmark experiments,Evt SNN is accelerated by 2.9-14.0 times compared to the event-driven framework EDHA.In the unsupervised training task of MNIST and IRIS data sets,Evt SNN is accelerated by 8.3 times and 8.32 times compared with EDHA,respectively.(3)The optimized simulation framework was transplanted to the embedded platform STM32L4R5 and Raspberry Pi Zero 2W,and the running speed and power consumption were calculated.On the Raspberry Pi platform,ANNs with different batch sizes are compared with SNNs.The experiments show that SNNs are suitable for unsupervised learning of real-time data.