Study On Artificial Neuron Based On Threshold Switching Device

Using the similarity between the behavioral characteristics of threshold switching devices and biological neurons,artificial neurons can be realized in a compact form,and then the human brain can be imitated to break through the limitations of traditional computer architecture.At the same time,biological neurons have the behavioral characteristics of random excitation,which is guidance for us to quantify the uncertainty of computing results.Nowadays,most researches only focus on a single device,while artificial neurons based on different device types have different behavioral characteristics and can be applied to different application scenarios,so systematic research is necessary.At the same time,the research on neuron probability characteristics is still very preliminary.The examples still demonstrated handwritten digit recognition which is a deterministic problem and the value of probability characteristics is not well clarified.In view of the above problems,this paper selects three different types of threshold switching devices for research.By using the same device structure and the same neuron circuit,this article systematically studied the characteristics of artificial neurons under different device types through circuit simulation,and summarized the relationship between the ability of artificial neurons to integrate information and the high resistance state of threshold switching devices.On this basis,this article studied the influence of the electrical parameters of the threshold switching devices on the characteristics of neurons by changing the input conditions,membrane capacitance and load resistance.The low-frequency excitation of the neuron and the power consumption of a single output pulse are determined by the high resistance and threshold voltage of the threshold switching device,and the high-frequency excitation of the neuron is determined by the switching time of the threshold switching device.Finally,in view of the probabilistic characteristics of neurons,a conductive bridge threshold switch device that meets the requirements of probabilistic neurons is selected for in-depth research.Furthermore,by constructing a probabilistic neural network to predict the non-deterministic breast cancer tumor diagnosis problem,a predictive result with quantifiable uncertainty is obtained.Compared with the traditional spiking neural network,the reduction degree of the original data probability distribution is improved by 81.2% in this study.