Design Of The Perceptron Neural Network Based On Memristor

The explosive growth of data calculations has brought severe challenges to the Von Neumann structure computer.Artificial neural network has powerful parallel distributed processing capability.At present,it has gradually become a research hotspot.Perceptron is a typical neural network,which is the origin of other networks.It is of great practical significance to study the hardware implementation of perceptron.Memristor is a kind of nonvolatile device with variable resistance.Its performance is similar to that of artificial neural network synapses.The memristor is thus an ideal device for building neural network circuits.According to the characteristics of neuron weight adjustment and memristor resistance change,single memristor perceptron neuron,reverse series memristive perceptron neuron and global-value memristive perceptron neuron are designed.Single-layer and multi-layer perceptron neural networks based on memristor are built with the global-value memristive perceptron neuron as the basic unit.Based on the performance similarity between memristor and neural network synaptic,single memristor perceptron neuron is designed.The neuron is composed of a single memristor weight module,an input weighting module,an information fusion module,a mapping output module and a feedback control module.The single memristor perceptron neuron adjusts its weights through the series structure of the memristor and the fixed value resistor,and then simulates the synaptic plasticity and memory characteristics of neurons.However,single memristor perceptron neuron has some problems,such as uneven weight adjustment rate,smallweight adjustment range and complicated operation process.Therefore,reverse series memristive perceptron neuron and global-value memristive perceptron neuron are proposed.The reverse series memristive perceptron neuron adjusts the weight through a pair of reverse series memristors.The global-value memristive perceptron neuron needs a memristor circuit,an absolute value circuit and a differential amplifier circuit to adjust the weight.Among them,the memristor circuit includes two pairs of reverse series memristors with different connection modes.The global-value memristive perceptron neuron has the best performance among the three memristor based neuron circuits.Its weight can be adjusted bidirectionally at a constant speed within the global-value range.Taking the global-value memristive perceptron neuron as the basic unit,single-layer and multi-layer memristive perceptron neural networks are built to solve the problem of multi-classification and linear inseparability.Through the logical OR operation example,the performance of a single global-value memristive perceptron neuron is verified,and the network accuracy rate reaches 100%.Through the example of judging the quadrant of the point in the two-dimensional plane,the multi-classification function of the single-layer perceptron neural network is verified,and the classification accuracy rate reaches 96.67%.Through the logical XOR operation example,the function of the multi-layer perceptron neural network to deal with linear inseparable problems is verified,and the network accuracy rate reaches 100%.Through an example of digital image recognition,the single-layer and multi-layer perceptron neural networks based on memotistor are compared from the perspectives of circuit structure,working principle and training efficiency,which further verifies the feasibility of the designed network.