Design Simulation Of Online Learning SNN Based On Synapses Of Spin Devices

With the development of artificial intelligence technology,Spiking Neuron Network(SNN)based on CMOS devices faces issues such as power consumption and reliability.Spintronic devices,such as the Hall-Bar,exhibit non-volatility and continuously tunable anomalous Hall resistance,which can emulate synaptic devices in neural morphologies.Furthermore,Hall-Bars have both computing and storage functions,enabling the capability of in memory computing,and therefore making them an ideal hardware unit for low-power neural morphology systems.However,current research on SNN based on Hall-Bar synapses is mainly limited to verify the synaptic characteristics of individual devices from a theoretical point of view.To further explore the potential of Hall-Bar synapses in SNN,we have done the following study:(1)Characterization of Hall-Bar synaptic properties.The Hall-Bar was tested by adjusting the numbers of current pulse with fixed amplitude and current pulses with varied amplitude.The test results show that the change trend of Hall-Bar anomalous Hall resistance is able to mimic the biological synaptic characteristics.A comparative analysis of the two pulse current methods was further conducted to determine the method of adjusting anomalous Hall resistance using the numbers of current pulses with fixed amplitude.(2)Hall-Bar model data testing and curve fitting.The anomalous Hall resistance of the Hall-Bar was investigated under different resistive states by adjusting the numbers of current pulses.The relevant parameters of the anomalous Hall resistance were determined by analyzing the test data,and a function curve was obtained by fitting the change of anomalous Hall resistance and over the numbers of current pulses to provide guidance for the design of the Hall-Bar model.(3)The first implementation of an SNN system based on Hall-Bar synapses using FPGA.Based on the relevant parameters and fitted curve of the anomalous Hall resistance,a single Hall-Bar was modeled using FPGA,and a Hall-Bar synapse array was designed to simulate the storage and updating of synaptic weights.In addition,modules for time-priority encoding,address encoding based on a polling arbitration mechanism,neuron processing kernel,and output layer were designed to provide the system with online learning capabilities.By configuring different system parameters and selecting different learning algorithms and neuron combinations,the learning effect of the system was explored.Finally,we implemented an SNN system with 10 neurons and 1,960 Hall-Bar synapses on an FPGA prototype verification platform.The comprehensive results show that the utilization rate of logic resources is less than 13%,and the utilization rate of BRAM resources is less than 5%.The system successfully performed online learning of the MNIST handwritten digit dataset,verifying the feasibility of the Hall-Bar synapse-based SNN system.Furthermore,the learning effects of different learning algorithms and neuron models were compared and analyzed.The results show that the combination of LIF neuron and STDP learning algorithm has the best learning effect.