Dynamic Recognition Of Spiking Neural Network Based On LiSiOx Memristor

On one hand,the development of computer science is limited by von neumann architecture and the failure of Moore's law.On the other hand,big data needs to process the extremely huge video data stream in an intelligent,real-time and efficient way by limited bandwidth and storage capacity,such as target motion pattern recognition,which requires the new generation of computing architecture and hardware.The neural networks and neuromorphic hardware process-in-memory shows a new path to solve this.The spiking neuron networks(SNN)we studied is a network with high biological plausibility based on brain-like research.It adopts a new computing architecture process-in-memory different from the current von neumann architecture.And it has the advantages of bionic brain such as ultra-low power consumption,intelligent learning and so on.Firstly,we fabricated the lithium silicate(LiSiOx)memristor as neuromorphic hardware,and realized the spike timing dependent plasticity(STDP)function of biological synapse on the basis of its memristor properties.Then,we proposed a single-layer unsupervised SNN,used dynamic vision system(DVS)sensor which inspired by the neurons in biological retinas with asynchronous signal transmission mode based on address event representation(AER)coding video as the input.Finally,we achieved dynamic recognition in videos based on the characteristic of LiSiOx memristor.In summary,(1)We successfully fabricated and characterized LiSiOx memristor.And an artificial synapse was designed based on its mechanism of memristor conduction.Its STDP function of biological synapse was also realized through the optimization design and modulation of neuron spike waveform.(2)We systematically studied the coding methods,learning rules and the neuron models of SNN for dynamic pattern recognition.We described how to use the sparse coding signals of DVS and the memristor to build a single-layer SNN.A pair of memristor synapses was designed to construct the synaptic array of the neural network for the two input signals of the motion pattern.We successfully identified twelve motion patterns by online and unsupervised learning.(3)We discussed in depth the problem of repeating learning and greedy learning in unsupervised network,introduced the learning capacity problem in the winner take all(WTA)neural network as a result of lateral inhibition delay.And we also studied its relationship with the range of initial distribution of the weight of the memristor synapse.The conclusion is that the initial distribution of the weights of the memristor synapses should be neither large nor small,which may lead to problems such as ineffective learning or repeatitive learning.(4)We focused on the relationship between initial weights of memristor synapses and learning capacity of neural network.By adjusting the forming voltage of LiSiOx memristor to optimize its initial weight distribution,the learning capacity of neural network is maximized successfully.A collaborative design method of software and hardware for dynamic pattern recognition is proposed based on this.