Study On Memristor-based Neural Network Hardware

In recent years,memristors as an emerging electronic component have important applications in both non-volatile memory and neurobionics.It is a device which has a capacitor-like structure,and it is capable to be write many times and be read nondestructively.Thus memristors are promising competitors for next-generation memory.Since the memristor is a two-terminal device and has an adjustable state,memristors have a resemblance to biological synapses,thus it is also of high research value in the field of neurobionics.Memristor-based large-scale neural networks are expected to become brain-like computing systems.Neural networks have been researched for decades.With the introduction of deep neural networks,application areas of neural networks have been greatly expanded.At the same time,neural networks are gradually reaching or even surpassing human levels of capability.Various smart devices based on deep neural networks are entering people's lives.However,neural networks also have some challenges such as high demands on computing resources and high energy consumption.Memristor-based neural networks are one of the possible ways to solve these problems.Currently,memristors are mainly used to build synapses in neural networks,while neurons built by memristors and neural networks implemented entirely by memristors have been less studied.The main difficulties encountered in the application of memristor-based neural networks contain the randomness of memristors,the accuracy degradation in analog operations due to sneak path and crosstalk in Crossbar array,large-scale integration of neurons and synapses,and compatibility with CMOS processes.This dissertation focuses on the memristors and their applications in neural networks,as well as,computing-in-memory neural network accelerator using memristors,the following works are carried out1.The memristor-based pulse-integral neurons and online learning synapses are modeled,designed and implemented,and their properties and performance are tested.The resistance of the memristor in the pulse-integral neuron represents the membrane potential.Thanks to the pulse frequency encoding,the randomness of the memristor has less effect on the neuron performance.The characteristics of the memristor-based neurons are statistically consistent with the neuron model.Because of the small area of memristorbased neurons,it is easier to integrate on large scale neural networks than capacitor-based neurons.It is suitable for building large-scale neural network systems.Online learning synapses convert the time difference between pre-synaptic signals and post-synaptic signals into amplitude-encoded pulses,and the online update of the synapse weight is realized by applying the pulse to the memristor.2.The Hopfield network and the multilayer feedforward network are simulated,built and tested.The memristor-based Hopfield network can update the neuron state synchronously or asynchronously,and it realizes a single-objective/multi-objective memory and confused memory.A multilayer feedforward network that supports online training is proposed.Based on the backpropagation algorithm,the superposition of the amplitudemodulated forward propagation signal and the delay-modulated backward propagation signal is applied to the threshold memristor,and the resistance of the memristor is adjusted according to these signals,thus the network is trained online.3.Two types of memristor-based spike neural networks are simulated,built and tested.A Winner-Takes-All network with short-term plasticity is proposed.Neurons form excitatory or inhibitory connections through synapses,the short-term plasticity is realized by the memristor conductance and capacitive memory circuit.This network shows the dynamic process of the ”winner takes all” model in the biological nervous system.A spikedbased convolutional neural network is proposed.The negative weights are removed during training,thus a network without negative weights is obtained.The absence of negative weights simplifies the design of synapse.It realizes the recognition of handwritten digits and achieves an accuracy of 97.1 %.4.A memristor-based computing-in-memory neural network accelerator is designed.The memristor in the Crossbar array stores 1 bit information,and multiply-accumulate operation is realized through the logic operation in the memristors and the counting of word lines by ripple counters.This method avoids the operation in the analog domain,and the influence of sneak path and crosstalk in the Crossbar array is reduced.The full-precision algorithm is supported by the accelerator.The energy consumption of computing-inmemory elements is only 318 m W.