Construction And Optimization Of The Threshold Device Based Spiking Neural Network

The traditional von Neumann architecture has a "storage wall" problem,which will result in significantly increased latency and power consumption when dealing with big data.Spiking neural network(SNN)simulates efficient information transmission between neurons in the real human brain,possessing the potential for big data processing and the advantage of low energy consumption.Neurons are a fundamental component of a spiking neural network.They could be constructed based on insulator-metal transition(IMT)threshold devices.However,the existing IMT neuron or device models presented problems such as poor universality,large model characteristic deviations,low simulation accuracy,and even poor model convergence,which limited the application of IMT neurons in large arrays.Consequently,the thesis proposed an optimized IMT device model,a neuron spiking circuit,and a general hardware IMT spiking neural network construction method.Firstly,based on the measured data of the fabricated IMT device and the IMT neuron circuit,a device model construction and optimization method based on measured data fitting was proposed.The scanning characteristics of the constructed IMT device model accurately reflected the actual IMT device performance.At the same time,the model was optimized based on the measured membrane potential data of the IMT neuron circuit.After optimization,the threshold voltage,hold voltage and activation frequency of the IMT neuron based on the proposed IMT device model coincided with the characteristics of the actual IMT neuron circuit,and the minimum activation frequency error between the model-based IMT neuron and the actual IMT neuron circuit was only 1%,which solved the problems of insufficient accuracy and lack of transient characteristics in existing models.Secondly,a neuron activation detection and spiking circuit was proposed considering the transmission requirements of neuron spikes.Based on the simplified threshold model of an ideal IMT device and the spiking circuit,the IMT neuron was constructed and tested,which generated standard spikes in a timely and accurate manner.The proposed spiking circuit provided powerful guarantee for inter layer signal matching and signal interaction between functional modules of large-scale multilayer spiking neural networks.A multi bit positive and negative weighted hardware IMT-SNN was constructed,and a computer-based weight training and mapping method was proposed.The recognition accuracy of the hardware network was tested on the MNIST handwritten numeral recognition dataset,which reached 79%.Furthermore,the correlation between input images and output neurons was analyzed,and the weight information of multi bit IMT-SNN was presented.The simulated recognition accuracy based on computer network was compared with the accuracy of the hardware IMT-SNN with limited weight accuracy.The hardware recognition accuracy was 10% lower than the simulated result.The constructed hardware network suppressed wrong prediction results,and no spikes of wrong categories were output in random numeral image test.Finally,based on the proposed IMT device model and spiking circuit,a multi bit weighted convolutional core circuit and a maximum pooling circuit were designed,and a hardware IMT hybrid network including the convolutional layer was constructed.The weight distribution of the convolutional layer and the full connection layer was analyzed to verify the strong feature extraction and inference capabilities of the network.The hardware hybrid network was tested using randomly selected handwritten numeral images,and the handwritten numeral recognition function of the network was verified.Based on the software network model and the mapped binary weight parameters,the hardware IMT hybrid neural network was simulated,which proved that the constructed IMT hybrid neural network achieved high recognition accuracy.The membrane potential characteristics and the recognition abilities between software and hardware hybrid networks were compared and analyzed.It was demonstrated that the proposed IMT neuron and the relevant spiking circuit were suitable for complex network construction,and the constructed IMT hybrid network had strong anti-interference ability.