A Strategy For Improving The Structure Of Deep Learning?Network

In recent years,a large number of excellent network models have emerged in the field of in-depth learning,which have achieved great success in image recognition,object tracking,medical image,automatic driving,bioinformatics recognition,data generation and other scenarios.Some models have achieved high accuracy,such as CapsuleNet[1][2],DenseNet[3],and others pay more attention to the speed of operation,such as MobileNet[4],SqueezeNet[5].At present,most of the research work focuses on the design of new models,and there seems to be little research on the general improved algorithm of existing models.Therefore,under the same task,how to improve the existing model so that both accuracy and speed can be taken into account has become the main research content of this paper.This paper attempts to start from three aspects:explaining the necessity of network structure improvement from the point of function fitting,interpreting and improving convolutional neural network from the point of differential characteristics,and understanding and improving selfencoder network from the point of manifold and mosaic.In the first part,in order to facilitate interpretation and application,the concept of tensor and vector-valued function is used to define deep learning mathematically,and the research on structural improvement of deep learning is transformed into the study of mathematical problems.The full-connected network and convolution network are unified in a set of frameworks,and the similarities and differences between them are explained.The significance of deep learning training is introduced.Then the fitting theory is used to solve the problem.Explaining the process of communication and training of in-depth learning shows that it is necessary to improve the network structure.In the second part,we explain what features are,how to propagate them forward,explain the reasons for the success of convolution networks,and point out the necessity of improving convolution kernels.Then we use truncated singular value decomposition to calculate truncated rank,which represents the essential characteristic number of the current layer,and observe the influence of different number of convolution kernels on the essential characteristic number and the subsequent generalization.With the change of precision and speed,an improved algorithm of convolution network structure based on truncated rank is proposed,which improves the network through iterative calculation until convergence.After that,this paper chooses resnet-34 network which is excellent in image classification task to improve the experiment,trying to find the relationship between the number of convolution kernels and network performance,and redesign the number of convolution kernels to make the network achieve the best.In the third part,for fully connected networks,we choose a representative model,self-encoder,to improve it.Firstly,we use the mosaic and mapping of manifolds in the background space to explain the working principle of deep learning,give the upper limit of network coding ability and the coding complexity of data,and then propose a space utilization ratio of quantified representation of the background space,and monitor manifolds by truncated rank.The utilization of the background space,and then the structure of the self-encoder to adjust the improved algorithm.Finally,the rationality of the algorithm is verified by experiments.deep learning,network structure,essential features,background space utilization.