Artificial Synapses,Neurons,And Neural Networks Based On Transition-metal-oxide Memristors

The rapid development of artificial neural networks(ANNs)paves the way to an intelligent society;memristors are essential for realizing ANNs due to their simple crossbar structure,and versatile performances for emulating synaptic and neuronal functions.However,a few problems still remain to be solved:at the material level,many transition metal oxides(TMOs)are mixed-ionic-electronic conductors,but some characteristics for ions are not yet well-understood;at the device level,it is quite difficult to modulate the memristive properties precisely,which limits the successful emulation of some neurological functions;at the system level,a new testing system for memristor-based ANNs is essential,because the computing systems based on memristors are different from current digital computing systems with the von Neumann architecture.By carefully selecting TMOs(Sr Ti O₃,WO₃ and Ta₂O₅)and mobile ions(_OV^··,H⁺and Ag⁺),materials with different ion mobilities are produced;by proper device structure design,memristors with non-volatile,partially volatile and fast volatile characteristics are realized;based on these characteristics,memristive synapses and neurons are obtained,and memristor-based ANNs are demonstrated in simulation and hardware.The main findings of this thesis are listed in the followings:1.Emulation of a bio-realistic synapse with the partially volatile Sr Ti O₃-based memristor.Pt/Sr Ti O₃/Nb-Sr Ti O₃ memristors,controlled by two state variables(oxygen vacancy distribution and diffusion coefficient),were fabricated by pulsed laser deposition and sputtering.By applying electric stimulations,the conductance state could be modulated continuously,and the conductance states were partially volatile in the absence of the electric field.Based on the characteristics of the resistive switching,the suppression triplet-STDP learning rule was faithfully demonstrated.The realized weight-dependent pair-and triplet-STDP learning rules are clearly in line with findings in biology.The physically realized triplet-STDP learning rule is powerful in developing direction and speed selectivity for complex pattern recognition and tracking tasks.2.Emulation of a bio-realistic neuron with the fast volatile WO₃-based memristor.Volatile memristors of W/WO₃/PEDOT:PSS/Pt with a battery effect were fabricated and investigated;the volatile resistive switching behavior and the battery effect were attributed to the proton injection at the heterojunction of WO₃/PEDOT:PSS and the migration of protons in WO₃.With a specially designed electrical circuit,quasi-Hodgkin-Huxley(quasi-HH)neurons with leaky integrate-and-fire(LIF)functions were physically demonstrated.Neuronal functions,including the temporal and spatial integration of input signals,local graded potentials with leaky features,and HH neuron-like spike firing were successfully realized.With multifunctions similar to their biological counterparts,quasi-HH neurons are advantageous over the reported HH and LIF neurons,demonstrating their potential for neuromorphic computing applications,such as realizing STDP learning rule in Pt/Sr Ti O₃/Nb-Sr Ti O₃ memristors.3.Simulation and optimization of memristive neural networks.By using non-volatile WO₃-based memristors as artificial synapses,a three-layer ANN was constructed to classify the MNIST handwritten digits.Based on the stochastic volatile memristor of Ag/Ta₂O₅:Ag/Pt,the dropout neuronal unit,which was probably dropped out in the training process,was demonstrated.The memristive neural network utilizing the dropout neuronal units showed a high accuracy in classifying handwritten digits by suppressing the overfitting in the case of insufficient number of training samples,and mitigating the error from the nonideality of the memristive synapses.The physical natures of the increased stochastics and accelerated volatility of the device were investigated by the kinetic Monte Carlo simulation,and attributed to the additive of Ag in the Ta₂O₅ layer.The implementation of the dropout neuronal units promotes the application of memristive neural networks in the face of limited training samples and the absence of ideal artificial synapses.4.Hardware demonstration of memristive neural networks.To implement neural networks in hardware,a testing system was designed based on FPGA and commercial chips on a printed circuit board(PCB),including DACs,ADCs,multiplexer,etc.Essential functions,e.g.generating pulses,measuring currents and capturing the neuronal responses,were all implemented in this system.By connecting the memristive arrays bonded with the PCB,the system successfully realized some demo tasks.Moreover,a training algorithm based on simplified STDP rules was proposed.With the ex-situ training for the memristive arrays,the system demonstrated the position recognition based on spatio-temporal information.In summary,we systematically study the way to realize artificial synapses and neurons with memristors,and discuss the construction of memristive neural networks in simulation and hardware.The work could be of great significance in developing ANNs with high speed and low power consumption.