Research On Evaluation And Optimization Of Bionic Degree Artificial Neural Networks

To build a more intelligent artificial neural network has always been the goal of researchers.This paper deeply studies the model of typical artificial neural network and its application,evaluating and optimizing artificial neural network from the perspective of bionics.Firstly,this paper compares and analyzes the application of several typical artificial neural networks,including KIII model,BP neural network,and LeNet5 model in pattern recognition(EEG recognition and face recognition).However,the whole performance of the model and the fit degree of running-in and real neural network is only by the level of the recognition rate,which is not enough in academia.Then,from the perspective of bionics,this paper constructs an index set to evaluate the bionics of artificial neural networks,and uses quantitative methods to analyze the bionics of the model as a integral analysis.In terms of qualitative analysis,the neuron equation,network structure and weight renewal principle of the model are compared and analyzed;In the quantitative aspect,the index set is constructed based on the bionic point of view i.e.small world,synchronization and chaos,which is analyzed on the model.The analysis results show that the Le Net5 model and the BP neural network have synchronization characteristics.But these two models still can't correctly reflect the real biological neural network.we find that the KIII model has certain small-world characteristics in structure and its network also exhibits synchronization characteristics and chaos characteristics,which is closer to the real biological neural network.Finally,a random edge adding algorithm is proposed to improve the performance of the convolutional neural network model based on the small world network idea.This algorithm takes the convolutional neuralnetwork model as a benchmark,and randomizes backwards and cross layer connections with probability p to form a new convolutional neural network model.The proposed algorithm achieves the goal of optimizing cross-layer connectivity by changing the topology of the convolutional neural network and provides new ideas for the improvement of the model.Based on the simulation experiment of fashion-minist and Cifar10 data set,we obtained the following results: 1)the model recognition accuracy and training convergence speed are greatly improved by random edge adding reconstructed models with aprobability p = 0.1;2)By the way of random edge adding with the number of edges is 15,the comprehensive performance of the model is optimal.