Analysis And Design Of Online Learning Circuits Of Memristor-based Neural Networks

In the field of neuromorphic computing that promotes the development of artificial intelligence,memristors are widely used for the inference acceleration of neural networks due to their non-volatility,low power consumption,in-memory computing and so on.But most existing memristor-based neural networks cannot take full advantage of memristors because they still use offline training.Therefore,online learning circuit designs of memristor-based neural networks attract more and more attentions because they can further accelerate the learning process and reduce power consumption and area of neural network circuits,which promote the applications of neural networks on mobile devices.First of all,two methods are proposed to modify existing memristor models for a novel memristor,which lays a solid foundation for the memristor-based circuit design and simulation.In addition,network circuits based on 3D memristor crossbars are designed for the inference acceleration of the capsule network.The convolutional operations and matrix-vector multiplications are realized with high parallelism,which improves the computation efficiency of the network.Then the capsule network circuit is applied to the digit recognition.Furthermore,in addition to existing memristor crossbars,a memory column,neuron block,convert block,and control block are designed for the online learning of fully-connected neural networks.The processes of forward propagation,backward propagation,and weight update in the network training are realized on the memristor-based circuits to reduce data transfer and conversion,accelerate the learning process,and apply to classification tasks.Moreover,in order to balance the high-precision requirements of synaptic weights and the non-ideality of memristors in traditional network training,a memristor-based synapse circuit is designed for the online training of hierarchical temporary memory networks.The online learning circuit designs of neural networks in this thesis have the advantages of simple synaptic structure,high efficiency of matrix-vector operations,parallel update of synaptic weights,and so on.They are expected to promote the online learning of memristor-based neural networks and further be applied to the neural network chip designs with high efficiency and low power consumption.