Study On 4T1M Synapse-based Memristor Neural Network With Its On-chip Learning And Logical Application

With the development of information society, neuromorphic plays an important role on automation, pattern recognition and artificial intelligence and neuromorphic is a technology that information processing of brain is simulated by large scale integrated circuit. Moore's law will soon reach its limit because of the physical limitation of CMOS process, which results in traditional synapse can not be further optimized in size and performance. It also directly hinders neuromorphic move toward the direction of high density and high reliability. As an emerging circuit element, memristor has been considered as an ideal candidate of novel synapse thanks to its unique advantages in high density, low power and especially nonvolatility, this is also provide an opportunity for the further development of neuromorphic.Firstly, this thesis makes a deep analysis on the theory of memristor and expounds the internal mechanism of HP memristor, including its mathematical model. In order to make HP memristor satisfy actual situation better, a HP threshold model and its detailed analysis of the memristance's change rules are proposed in this thesis. Secondly, two existing classical memristor bridge synapse circuits with their corresponding neural network architectures are analysed in detail and verified by the simulation results. Then, a novel 4T1 M synapse composed of a memristor and four transistors is designed and the simulation results verify that it can function as positive, zero or negative weight. Moreover, circuit schematic of a4T1 M synapse-based neuron is designed and this thesis also describes the implementation principle of input signals' weighted summation by mathematical analysis. Furthermore, a4T1 M synapse-based multilayer neural network is also designed. All the analysis results of these designs are validated by the simulation results. Then, a 4T1 M neuron-based circuit to implement the learning of AND-function based on the error back-propagation algorithm is designed. Finally, this thesis presents some related logical applications of 4T1 M synapse-based neural networks, including the design methods of linear classification with single layer neural network and nonlinear classification with multilayer neural network, the crossbar array of 4T1 M synapse and the application of configurable logic with the 4T1 M synapse-based crossbar array.The researches of this thesis provide a new idea of the development of memristor neural networks. It will improve the traditional information processing model and make it possible for neuromorphic move toward the direction of high density and high reliability.