The Study On Various Memristive Neural Network With STDP Rules

Artificial neural network, shorted for neural network, is a mathematical model which is abstracted from the function and structure of biological neural networks. The artificial neural network is an adaptive, nonlinear processing system which is based on the results of neuroscience research. And the system uses the interaction of large number processing units to achieve brain-like information processing. Because of the high complexity and density of human brain, the artificial neural networks will develop to be complex and large-scale in the future. To achieve this goal, the important components—neurons and synapses of neural networks should have the nanometer size and low power consumption characteristics. Due to synaptic plasticity, dynamic memory and storing ability, therefore, the choice of synapse for artificial neural network is really the crucial point. However, the development of memristor has brought hope to solve this problem.Memristor is believed to be the forth fundamental two terminal passive circuit elements, besides the resistor, the capacitor and the inductor. The existence of such element was postulated in theory by Leon Chua in 1971, based on logics and symmetry. Its memristance changes with the amount of charge and flux passing through it. In 2008, HP Labs has realized an actual memristor on the nanoscale level for the first time in the history. Since then, a lot of researchers have focused their attention on the memristor. Because of its unique non-volatile memory and nanoscale characteristics, it has prospective promising applications in nonvolatile memory, artificial networks, pattern recognition and signal processing. All these make memristor ideally suited to act as synapses in artificial neural networks.This thesis describes several memristor models, and analysis the theory and characteristic of memristor through a mathematical way and simulation. Two mathematical models of the STDP learning rule are also introduced. This thesis uses the asymmetric STDP as the learning rule of memristive neural network, combined the genetic algorithm which include adaptive mutation and topological variation, to complete the comparison of different memristor models’ learning performance. For the lack of HEBB learning rule in Pavlov experiment, this thesis proposes a forgetting memristive neural network with the symmetric STDP learning rule to verify the associative learning, correcting and forgetting ability of the memristive neural network, which is much closer to the biological system. This also provides a basis for the research of more intelligent bionic neural network in the future.