Artificial neural network(ANN) is a nonlinear information processing system which can simulate the working mechanism of the physiological neural activities of the human brain, ANN has the abilities of self-organizing learning, parallel processing and fault tolerance. Which can be applied widely in the field of data analysis, artificial intelligence, prediction and intelligent information management.As the essential part of the study of artificial neural network, the implementation technology of ANN can be divided into hardware implementation and software implementation. ANN with software implementation has the characteristics of low parallel processing speed, which is difficult to meet the requirements of neural network information processing of the real-time. However, The neural network with hardware implementation shows the rapidity and accuracy of the information parallel processing,which is very beneficial in the complex data processing. Compared with digital circuit implementation, the analog circuit implementation has the advantages of small chip area,high processing speed, good linearity mean while has the disadvantages of easy disturbed by outside noise, Which is the inevitable trend of development of the research of the implementation technology of ANN.There are not many types of analog circuit implementation of the ANN and the performance of the existing ones can be improved at present. Based on a comprehensive study of the hardware implementation of the ANN, The paper focusing on the design of CMOS analog circuit implementation of the single layered perceptron neural network and the fuzzy neural network, and the application results of the proposed ANN were realized by simulation and verification. The main contents and innovations of this paper are as follows:(1) In order to realize the parameters adjustment and performance improvement of the neural activation function circuit, a trapezoidal activation function circuit, a sigmoid activation function circuit and a gaussian function circuit was designed in this paper.The proposed gaussian function circuit can able to adjust the parameters of the output function through the tail current and the bias voltage which is only comprised of two cross coupled differential pair, and can increase the outputs number of the circuit only by adding a tail current source and a differential pair. The sigmoid function circuit can realize the unipolar and the ambipolar sigmoid functions output. The amplitude, the gainfactor and the threshold of the generated function can be adjusted through external bias voltage or current.(2) In order to realize the adjustment of the weight value of the synaptic circuit automatically, a current-mode four-quadrant multiplier circuit and a linear adjustable operational transconductance amplifier was proposed. The weight of the proposed synaptic circuits is adjustable by changing the external bias current and the circuits has the merits of good linearity, high precision and wide input range, which can be applicable to the neural network as linearly weighted circuit.(3) In order to solve the problems of the complicated structure and low accuracy of the hardware implementation of the fuzzy neural network, a gaussian function circuit, a minimum and maximum circuit and a current-mode multiplier / divider was designed.Which can be used to simulate the fuzzifier circuit, inference engine and the defuzzification process of the fuzzy neural network. The proposed circuits has the merits of simple structure, high precision and good expansibility. Finally, the hardware implementation of the fuzzy neural network was presented in the paper and the result realized by simulation and verification.(4) Based on the neuron synapse circuits and the activation function circuits, the analog circuit implementation of two feed-forward artificial neural network was desi gned and verified for the XOR problem and data classification. |