Implementations Of Spiking Neural Networks With Electronic Synaptic Devices

Today's density-driven development of memory technology gradually reaches its physical limits,and functional-driven cognitive memory devices can become successors to the continuation of Moore's Law.In recent years,the study of neural morphology and cognitive computing has aroused widespread interest.In the study of the development of bio-inspired computing as the next generation(later Moore,non-von Neumann),we have discussed different new nonvolatile resistivities in various applications of potential applications in various fields Memory technology,including(i)phase change memory(PCM),(ii)memristor,for building neural form hardware.We focus on the use of new nonvolatile memory devices to achieve biological synapses long-term potential(LTP),long-term depression(LTD)and spike-timing-dependent plasticity(STDP).We have developed new programming methods and simplified STDP our study shows a large-scale implementation of effective neuromorphic systems with two different pulse neural network(i)supervised learning(ii)unsupervised learning.Synaptic ideal alternative to the principle of neural networks in STDP model uses experimentally measured mathematical model of the process by adjusting the learning rule,adapt the physical device having asymmetry,nonlinear problems and inconsistencies and based on learning outcomes,for further improvement of requirements for electronic synapse device.In summary,(1)We discuss how to use phase-change memory(PCM)techniques to simulate biological synapses in large-scale neural morphological systems that are easy to implement,using the thermal accumulation/diffusion effects of phase change materials to achieve the use of simple rectangular pulses Implement STDP function.(2)We discuss how to use the memristor to realize the STDP function of biological synapses,measure the STDP curve of electronic synaptic devices under different resistance states,and construct the electronic synaptic model which can express the actual work.(3)We describe how to use the electronic synaptic device to construct a supervised learning spiking neural network,using the ReSuMe model.By adding the refractory period,the problem of learning failure caused by the asymmetry of the electronic synaptic device SET/RESET is solved.The balance of the precision of the device resistance control and the complexity of the pulse waveform is discussed,and the robustness of the system is studied.We propose a scheme for constructing the neural network in an electronic synapse crossbar array.(4)We discuss how to construct an unsupervised learning pulse neural network using electronic synaptic devices,and propose a scheme to improve the accuracy of electronic synaptic devices.We study how to quantify the stability of neural networks and discuss how to optimize electrons Synaptic devices to accommodate neural networks.