Full Circuit Design Of Memristive Neural Network And Its Applications

In recent years,artificial neural network has become the popular research direction in the field of artificial intelligence,and has successfully solved many practical problems in the fields of robotics,industry,biology,medicine and economics.With the advent of the internet and the era of big data,the information to be processed by neural networks has presented explosive growth.Computers based on von-neumann structures can hardly cope with the increasing demands of data processing.Therefore,it is urgent to establish neural network hardware circuits that can provide real-time data processing and super-large parallel computing.Memristor crossbar array,which has the inherent advantages of parallel computation and integration of memory and computation,provides a new solution to this problem.The memristor-based neural network circuits have been widely concerned by researchers.After years of development,the research of memristive neural network has yielded fruitful results.But there are two problems need paying more attention,(1)the training process of the network is usually realized through the offline system.Computer or digital chip is needed for auxiliary calculation,(2)most of the existing designs do not give a complete circuit of neural network,can not perform a complete online learning process on the hardware.To address the above problems,this dissertation focuses on the full circuit design of memristive neural network.According to the actual memristor data,the corresponding memristor model was constructed,and then the memristor synaptic circuit and neuron circuit were designed.The algorithm was introduced into the memristor circuit and the memristor learning circuits were designed.Finally,we designed the full circuit implementation of memristive neural network according to the basic theory of neural network.Some practical problems such as image recognition and curve fitting can be solved by all-hardware online learning system.The main work and achievements are summarized as follows:(1)Considering the diversity of actual memristors,a drift speed adaptive memristor(DSAM)model is proposed.The proposed model can match the i-v characteristic curves and memristance variation curves of various physical Memristive devices with an averageaccuracy of large than 94.5%.The DSAM model is effective for continuous excitation and pulse excitation,and can carry out various fine fitting of the real memristor data by using adjustable parameters.(2)Novel circuits based on memristors for implementing electronic synapse and artificial neuron with current and voltage modes are designed.Two kinds of reverse-connected memristor synaptic structures are designed to realize current and voltage multiplications of synaptic weight calculation.Positive,zero and negative synaptic weights can be obtained by adjusting the memristance.A variety of memristor neuron circuits are designed based on the CMOS-memristor hybrid integration technology.This structure can realize parallel programming,and can be applied to large-scale neural network integration.(3)We faithfully replicate Froemke's triplet-STDP rule using SrTiO3(STO)-based second-order memristors experimentally and numerically.Moreover,a LIF neuron and triplet-STDP Learning circuit were designed.Based on the designed neuron,the temporal and spatial integration of excitatory postsynaptic potentials,postsynaptic impulse generation and backpropagation,synaptic weight modification,according to the suppression triplet-STDP rule,were vividly realized.(4)In order to solve the problem of initializing memristor logic circuits,a universal memristor logic circuit with self-learning function is proposed,that is,different memristor states can be obtained automatically through self-learning without initializing.Moreover,the proposed design offers a uniform circuit that is configurable to perform various types of logic,including Boolean,IMPLY,and random logical combinations.Compared with the previous logic circuit,the proposed circuit is a more general design,which can handle different logic operations without initialization,simplify the operation steps and speed up the logical operation.(5)To solve the problem of the full circuit of memristor neural network,the implementation of LMS learning algorithm is introduced into the design of neural circuits,and two kinds of memristor self-learning neural circuits are established.Extending on the designed neuron circuit,we proposed the full circuit implementations of the Perceptron Network,RBF Neural network and multi-layer neural network using hybrid CMOS-memristor integrated.Using image recognition,curve fitting and XOR operation ofan application level validation of the proposed circuits were successfully demonstrated.Compared with the existing memristor neural network circuits,the outstanding advantage of this design is that all the operations are completed on the circuit,without the need for computer-assisted operations,and the integration of memory and computation is truly realized,greatly improving the speed of operation.In addition,the implementation of online learning in the circuit can update the memristor weight in real time according to the changes of environment and demand,improving the system's tolerance to memristor variables and noise,and enhance the robustness and accuracy of the application.(6)In order to perform real-time and efficient image restoration task,a memristor-based continuous Hopfield neural network circuit is proposed.The designed HNN circuit can automatically implement image restoration through self-organizing network operations,opening an opportunity to transfer the training from the software to on-chip implementation.This can speed up the processing speed of image restoration task.In addition to the speed advantage,our experiment also demonstrated that the image restoration achieved by the circuit has higher accuracy,and has a good tolerance for noise and memristor variables.The research results of this dissertation can provide a new idea and reference for the hardware implementation of online learning algorithm and memristive neural network,and make foundation for hardware implementation of larger memristive neural network circuits.In addition,the full-hardware implementation of memristive neural network can accelerate the operation of neural network,which provide a hardware accelerated platform for the big data era,and bring new hope for realizing more complex applications in the field of artificial intelligence.