The Design And Research With Two Types Of Neural Network By CMOS Analog Circuit

Neural networks are not only intelligent systems to simulate the basic characteristics of the human brain, but also an information processing science. Neural networks have the characteristics of adaptive learning, nonlinear mapping, distribute parallel processing. Neural networks simulate the brain structure from individual neurons to information processing function. Neural networks are widely used in the field of nonlinear systems, network troubleshooting, aerospace, intelligent robots.The study of the neural networks is divided into three areas: theoretical research,applied research and implementation of technology research. There are two methods to implement neural network, software implementation and hardware implementation.There is low processing speed, the degree of parallelism and low defects by software.So it is difficult to meet the requirements of neural network information processing of the real-time. However, Hardware implementation can overcome these shortcomings,which is very beneficial in the complex data processing. Therefore, the hardware implementation is the inevitable trend of development of neural network. Because of analog CMOS circuit to implement neural network has a simple structure, integrated high speed, small footprint chip area, high integration, low power consumption, this paper use analog CMOS integrated circuit to implement neural networks.Neural network model has a representative are: BP back propagation network,radial basis function RBF networks, ad hoc networks, sensor feedback Hopfield network, CMAC CMAC networks, fuzzy neural networks and so on. Now, there are some neural networks are implemented by hardware, such as BP back propagation network, radial basis function RBF network, perception, etc., Based on a comprehensive study of the neural network, this paper studies self-organizing competitive neural and fuzzy neural networks. Around these two neural networks, the paper made the following related work:(1) The weight of neural network model cannot adjustable. This paper design a linear adjustable operational transconductance amplifier and a current multiplier circuit as a synaptic circuit, the weight is adjustable by changing the external current function, and a simple circuit structure design, high linearity. So the designed circuit can be used in the neuron circuit.(2) Based on self-organizing competitive neural network algorithm competitive layer is difficult to implement, this paper designed a current-mode circuit to implement competitive by comparing the size of the current. The circuit can be easy to implement and integration, together with the neuron model can implement a self-organizing competitive neural network.(3) To solve the problem of fuzzy neural network complexity low accuracy,Gaussian function circuit, seeking a small circuit, fuzzy circuit design was optimized.So the whole circuit can improve the accuracy and speed of the fuzzy neural network.Finally, the design of fuzzy neural network can approach a nonlinear function. The result realized by simulation and verification.