Sensitivity Analysis On Second Type Rational Spline Weight Function Neural Network With Cubic Numerator And Linear Denominator And Its Application

The traditional gradient descent neural network has some disadvantages, such as slow convergence, local minimum and so on. Although many scholars have improved it from many aspects, these shortcomings can't be solved fundamentally. So in this paper, using cubic numerator and linear denominator rational spline as weight function, construct second type spline weight function neural network.This spline weight function neural network overcomes the shortcomings of slow training speed, local minimum and before the network training starting, the dimensions of input and output can be determined according to the number of attributes of sample data and levels taken into. Cubic numerator and linear denominator rational spline is used to replace the constant weight as the weight function related to the training sample, so it can reflect the information of training samples. The output of the neural network will change when the input sample or the weight is disturbed. It is very important to analyze the influence of these changes on the whole network system in theory. Sensitivity is an important index of neural network, and it can be used to analyze the impact of these changes on the system. Firstly, this paper combines cubic numerator and linear denominator rational spline with spline weight function Neural Network of second type and studys its learning algorithm. Then the sensitivity on second type rational spline weight function Neural Network with cubic numerator and linear denominator is analysised using the statistical methods. And then, Matlab simulation experiment is carried out which compare the learning ability, generalization ability of the traditional gradient descent neural network with the Neural Network in this paper.In the end, second type rational spline weight function Neural Network with cubic numerator and linear denominator is used to design neural network special chip. Neural network dedicated chip, with high speed, high real-time characteristics, suitable for speed and real-time requirements of the occasion. In this paper, the FPGA realization of second type rational spline weight function Neural Network with cubic numerator and linear denominator has been designed, and are simulated on the quartus and Modelsim, It can be seen from the experiment that this kind of FPGA implementation of this spline weight function neural network has strong generalization ability, suitable for the designed neural network chip.