The Appilications Of Memristive Neural Networks

Because of the memristor’s passivity, no consuming power, memory andnanometer scale,the memristor will have a breakthrough in the field of artificialintelligence, especially in the artificial neural network theory and application.Previousneural network circuit design is limited to the design of weight, but the appearance ofmemristor break the limitation and researchers begin to research the theory ofmemristive neural network. The memristor include linear memristor and nonlinearmemristor. The linear memristor has two memristor values, but the nonlinear memristorhas continuous dynamic countless memristor value. This article propose two memristiveneural network model based on the linear memristor and nonlinear memristor.The linear memristive neural network model can be applied to xor operation, lineardiscriminant, image edge detection and so on. The model can solve many easy questions.This article solves the XOR operation with the model. The model include three inputlayer, hidden layer and output layer, and the design of input pulse, hidden layerexpecting pulse, output layer expecting pulse,learning rule and feedback voltage isgiven. The input layer is xor boolean table. The expecting pulse is the expecting outputof the hidden layer and output layer. The feedback voltage is the realization of thelearning rule and the effect of the feedback voltage is changing the memristor value.The memristors are programmed repeatedly by input pulse and expecting pulse so thatthe model can achieve the requirements. Theory and simulation results demonstrate themodel can be applied to xor operation. The model has a engineering property is thatweight(memristor value) can be stored at last when powered off.The nonlinear memristive neural network model can be applied to patternrecognition, traveling Salesman, the robot finding its way and so on. The modelcompute parallel and very fast in engineering. This model will be applied to patternrecognition in the article. The pattern recognition is mainly to identify different people’shandwritten signature. The model is divided into input layer and output layer. Thedesign of the timing pulse, input pulse, expecting pulse, feedback voltage, learning ruleand so on is given in the article. Before training the model, the features are extracted bythe signatures and should normalize to specific voltage range. So the input pulse is thenormalized features. The timing pulse is necessary in the circuit which decides thecircuit’s execution time. The timing pulse cycle has the connection with the training samples (signatures). The expecting pulse judge whether the training samples arecorrect classification. The timing pulse, input pulse and expecting pulse decide thefeedback voltage value, and then the feedback voltage will program the memristor sothat the model can identify the signatures. Training samples and non training sampleswill be tested by the model. Theory and simulation results demonstrate the model can beapplied to pattern recognition. The model has the engineering property as follows: First,restore last weight (memristor) after power off. Second, the model not only trains thesamples to program the memristors, but also can be identify. Third, self-adapt ability.Because of restoring the last weight after power off, the model is more easier to trainnew samples, at the same time, the neurons input layer and output layer can beincreased or decreased more easily.