Study Of Silicon Photonic Neural Network Model

With the coming of big data age,the amount of data increases rapidly.The deep learning model for large-scale data processing based on neural network will be more and more complex.The von Neumann architecture applied in traditional computers will not meet the increasing demand for computing.Therefore,how to build reasonable software and hardware platform and upgrade the calculation performance of the neural network in speed and power consumption is in urgent need.Fortunately,photonic neural network based on photonic devices has the advantages of ultrafast speed,large bandwidth and low power consumption,and it's an ideal choice to achieve large-scale information processing tasks in the future.The photonic spiking neural network studied in this thesis is one of the photonic neural network.For the photonic spiking neural network,the thesis has completed the following three parts:(1)Inspired by biological leaky integrate-and-fire(LIF)neuron,we designed laser LIF neuron based on two-section excitable vertical cavity surface emitting laser.The designed laser LIF neuron had at least 7 orders of magnitude of improvement in terms of computing speed and at least 10 times reduction in power consumption when comparing to biological LIF neuron.To have a better understanding of the characteristics of the designed laser LIF neuron,we have thoroughly investigated its DC response,single pulse response,decision latency and double pulse response.And the comprehensive grasp of the excitability of the laser LIF neuron was good foundation for the studying of photonic spiking neural network.(2)We have studied the temporal and spatial information of the spikes and the encoding and decoding schemes based on the spikes.Considering the information coding capacity of each scheme and the difficulty when realizing it,we choosed binary coding scheme as the study object.And we firstly designed logical AND gate,logical OR gate and logical NOT gate based on single laser LIF neuron,and recurrent binary adder and binary multiplier based on several laser LIF neurons.And their principles were all verified.(3)According to the structure and function,three different physical and mathematical models of photonic spiking neuron and two different photonic neural network architecture(photonic feed-forward spiking neural network and photonic recurrent spiking neural network)were defined.In the application level,we proposed real-time radio frequency(RF)information processing architecture using the photonic spiking neural network.In the scheme,we utilized optical pulse sampling technique as the bridge between the RF front-end system and photonic spiking neural network.And the optical pulse sampling technique accomplished pre-coding of RF information at the same time.Finally,the advantages and disadvantages of the thesis were summarized and the future development direction of photonic spiking neural network has been forecasted.And I sincerely hoped that this thesis could be used as reference and guidance for future scholars.