Theoretical Research And Application Of Perceptron With Multi-pulse Type Activation Function

As one of the hottest research topics in recent years,the wave of deep learning has swept all walks of life with great brilliance.From simple handwritten number recognition and text classification,to style transfer and text generation,to future unmanned driving and intelligent robots,deep neural networks are inseparable.The perceptron,the basic unit of deep neural networks,is the basis of these complex models.The perceptron model has been studied for a long time,but traditional perceptrons can only solve simple linearly separable problems,and even the XOR problem cannot be solved.Therefore,according to the shortcomings of the perceptron model,this paper proposes a perceptron model and its training algorithm based on the multi-segment pulse activation function.This perceptron can perform high efficient classification for piecewise linear separability and the training algorithm can find the weight for this perceptron.The main innovations achieved are as follows:(1)There are many researches on solving algorithms for nonlinear separable problems,but there is no efficient algorithm for the special nonlinear separable problems such as piecewise linear separable problems at present.Therefore,the piecewise linear separability problem is first proposed and defined in this paper,and a targeted multi-domain pulse activation function perceptron model is proposed.Compared with the traditional perceptron,this model can make full use of the piecewise linear separability of the data set,and can carry out effective classification while maintaining its simplicity.(2)The activation function of the multi-domain impulse activation function perceptron for different practical problems may be different,and the original data may have very high dimensions,which further makes the definition of the activation function of the model more complicated.Therefore,this paper proposes to aggregate the data set by Kmeans clustering algorithm first,so as to obtain the appropriate number of clusters.Then,the boundary of activation function can be obtained by solving the optimization problem,and then the multi-domain pulse-type activation function perceptron model can be determined.(3)Once the boundary of the multi-domain activation function perceptron is determined,only the weight vector of the perceptron needs to be further determined.In order to solve the weight vector,this paper proposes a specific training algorithm of multi-domain pulse activation function perceptron,which will automatically converge when the data set is piecewise and linearly separable,and obtain the weight of global convergence.In short,this article mainly improves the traditional perceptron model and proposes a perceptron learning algorithm suitable for piecewise linear separable,which has achieved good results in some scenarios.However,the wider scope of application of this model needs further verification.