Simulations of Artificial Neural Network with Memristive Devices

The memristor has been hypothesized to exist as the missing fourth basic circuit element since 1971 [1]. A memristive device is a new type of electrical device that behaves like a resistor, but can change and remember its internal resistance. This behavior makes memristive devices ideal for use as network weights, which will need to be adjusted as the network tries to acquire correct outputs through a learning process. Recent development of physical memristive-like devices has led to an interest in developing artificial neural networks with memristors. In this thesis, a circuit for a single node network is designed to be re-configured into linearly separable problems: AND, NAND, OR, and NOR. This was done with fixed weight resistors, programming the memristive devices to pre-specified values, and finally learning of the resistances through the Madaline Rule II procedure. A network with multiple layers is able to solve difficult problems or recognize more complex patterns. To illustrate this, the XOR problem has been used as a benchmark for the multilayer neural network circuit. The circuit was designed and learning of the weight values was successfully shown