Study On Photonic Spiking Neural Network Learning Algorithm

At present,artificial intelligence(AI)has attracted more and more attention,and the application scenarios of artificial neural networks have become more and more extensive.However,the power consumption of operations during neural network training and inferencing is very high,especially in the face of more complex requirements and tasks.Therefore,the hardware architecture and software algorithm that can realize the high-speed and low-power operation of the neural network are particularly important.Spiking neural network(SNN)is known as the third generation of neural networks,with excellent power consumption efficiency.The spike timing-dependent plasticity(STDP)biological mechanism is the brain-like neuromorphic architecture.Light has the advantages of high speed,high bandwidth,low power consumption,and low crosstalk in propagation.Therefore,the photonic neuromorphic system that integrates photonics technology and the non-Von Neumann architecture of neuromorphological computing is one of the best hardware architectures that realize high-speed and low-power operation of neural networks.Based on the concept of collaborative design of optical hardware structure and SNN learning algorithm,this thesis is dedicated to the study on photonic spiking SNN learning algorithm.The content is summarized as follows:1.For the first time,a simple optical hardware structure based on a single photonic neuron device to realize exclusive OR(XOR)operation in a single step is proposed.In this scheme,through injecting dual polarized optical pulse into a single vertical cavity surface emitting laser with a saturable absorber(VCSEL-SA)photonic neuron,the XOR operation can be realized in one step due to the polarization mode competition effect.The results show that the XOR operation can also be accurately achieved for high-speed serial spike trains,and it is robust to input noise and time jitter.2.We propose a single-layer photonic SNN network structure and a hardware-friendly supervised learning algorithm for supervised learning and classification tasks.Based on VCSEL-SA and vertical cavity semiconductor optical amplifier(VCSOA),a self-consistent learning model is designed,which can realize a complete learning and inferencing.The simulation results show that the network can achieve spike sequence learning and spatiotemporal pattern classification tasks.3.We proposed the network structure of multi-layer photonic SNN based on VCSEL-SA,and designed a hardware-friendly multi-layer supervised learning algorithm.VCSEL-SA has both X-polarization(XP)and Y-polarization(YP)modes.The synaptic connection between the XP and XP(YP)of two photonic neurons is used to achieve excitatory(inhibitory)connections.Based on VCSEL-SA,a self-consistent calculation model of multi-layer photonic SNN is derived,Besides,the XOR operation is realized,and other logical operations can also be realized.