The Research And Implementation Of A Variety Of Feed-forward Artificial Neural Networks Hardware Based On CMOS Analog Technology

Artificial neural network is an artificial intelligence system which can simulatebiological neural network with distributed parallel processing, nonlinear mapping,adaptive learning and fault tolerance, and other functions. There are two main types ofartificial neural network method, software and hardware implementation. Theartificial neural network realized by software method is hard to meet the requirementsof real-time because of its shortcomings of low degree of parallel and slow processingspeed. These disadvantages make theory research disconnected with practicalapplication. However, the artificial neural network implemented with hardwaremethod overcomes the shortcomings of the former. It can process the signal massively,especially complex data and meet demands of practical application.In integrated circuit design, it is difficult to design an analog circuit, becausethere is a compromise between power, speed, gain, precision and area, in addition, theeffects of the layout on analog circuit is greater then digital circuit. It has brought newchallenge to the design of analog circuit. Considering current-mode analog circuit hasthe advantages of fast response for input signal variation, good linearity, low powerconsumption, etc, the feed forward artificial neural network in this paper is achievedby current-mode analog circuit.The work focuses mainly on the design of analog circuit implementation of twofeed forward artificial neural networks, namely, single layered perceptron neuralnetworkand radial basis function neural network and their applications are alsointroduced:(1)Elaborate the research background and meaning of hardware implementationof artificial neural network, and discuss the present research situation anddevelopment directions at home and abroad.(2)Introduce biological neural model, and discuss the structure and principle ofSLPNN and RBFNN.(3)Using TMSC0.35μm CMOS process, a current-mode linearly classifier whichcan classify linearly non-separable data is designed. The weighting coefficients can beobtained based on Fisher’s linear discriminant analysis by MATLAB software. Theproposed circuit is simulated and analyzed by PSPICE.(4)The block circuits of RBF neuron are designed. A two-input/one-output RBFneural network with two hidden nodes composed of these circuits is implemented and verified for the XOR problem by PSPICE. The network parameters are set by newhybrid shuffled frog leaping (NHSFL) algorithm.(5)The layout basic processes, tools types, analog layout problems and solutionsare introduced, and the layouts of each major cell circuits, linearly classifier and RBFneuron circuit are also finished.