Research On Small World Topology Based Neural Network

In recent years,deep learning has begun to play an increasingly important role in computer vision,natural language processing and other fields.At the same time,convolutional neural networks are considered to be one of the most classic and powerful algorithms in the field of deep learning,especially computer vision.The popular convolutional neural network models mainly include ResNet,DenseNet,AlexNet,VGG,etc.But in the field of image recognition,it takes a lot of attempts to design a network model that meets the characteristics of the image,which not only includes the model's parameter adjustment and depth Deepen,but also include the design of the network model.With the development of science and technology,people have found that in different fields such as technology,transportation,social relations,the network is not a simple rule or random network,especially in the internal structure and functional connection mode of the human brain.They are not similar to the rule network.With high local clustering and short average path length properties similar to random networks,small world networks have been widely studied and discussed in recent years as a typical complex network model with short average path lengths and large clustering coefficients.Therefore,we speculate that relying solely on mathematical methods to adjust the parameters between levels may cause the artificial neural network and the biological neural network to deviate further in structure,and cannot reflect the structural characteristics of the real brain neural network.Based on this,this article starts from the brain-like small world network topology,directly improves the structure of the neural network,proposes a new type of neural network connection method,and designs a neural network model with small world characteristics.Only consider the improvement from a mathematical point of view,abandon the traditional network hierarchical structure,and randomly generate a small-world network topology without a hierarchical structure as a network training model according to different probabilities.At the same time,in order to analyze the effect of different attributes of the small-world feature map on the network performance Impact,on the CIFAR-10 training set,multiple sets of experiments were designed for network performance analysis and verification.By comparing the small-world neural network under different parameter configurations,the necessity of the above improvement is verified.The training results show that the improved model in this paper has a faster convergence rate and higher approximation accuracy than today's traditional convolutional neural network.The stability is also stronger.