Research On Parameter Selection And Visualization Based On Convolutional Neural Network

With the advent of the big data era,the Convolutional Neural Network(CNN)has become a research hotspot in the field of deep learning because of its superior performance in image classification.However,the deep learning model is based on a large amount of data training and often regarded as "Black box" lacking reasonable explanation for model decision.Therefore,this paper will analyze CNN in detail from the theoretical,experimental verification and visualization research,then obtain the factors affecting the performance of CNN,and realize the model explained.Because the structural design and parameter selection of convolutional neural network model have great influence on image classification effect and model convergence efficiency,this paper obtains optimal activation function,gradient descent algorithm and batch normalization through a large number of comparative experiments and numerical analysis.Optimize the selection of parameters in the model structure and apply it to the typical two types of data sets.At the same time,Tensorboard is used to visualize the CNN model structure and parameter change process,to provide theoretical support for optimizing the specific parameters of the model in practical applications,and to assist the parameter adjustment work to obtain the optimal model.An improved CNN structure model is proposed for the MNIST dataset.The pooling layer is replaced by the convolutional layer to obtain a 9-layer network structure model.The accuracy of the model classification is improved by setting optimal parameters.For the Cifar-10 dataset,the model is optimized from the simplest convolutional neural network through the parameter adjustment process.The experimental results show that increasing the number of network layers can improve the accuracy,but the number of layers is not necessarily to make the model better.Gradient in the deeper network will gradually disappear in the back propagation.Therefore,a residual network is proposed to solve this problem.The accuracy of the model classification is eventually improved to91.55% and the model has a strong generalization ability,realizing a significant improvement in the accuracy of the model for image classification.In order to understand the working mechanism and prediction principle of CNN model and realize the interpretability of deep learning model,this paper uses deconvolution network to visually analyze the image features extracted by convolutional layer,and proposes class-based activation mapping.The results show that the shallow convolution layer mainly learns the simple structural information of the image,while the deep convolutional layer learns the content is more complex and abstract.The content of deep convolutional layer learning is a summary of shallowsimple information,which plays an important role in classification tasks.In addition,the combination of the gradient-based guided backpropagation method and the deep convolutional neural network can locate the target more clearly and achieve high-resolution visualization based on a specific category.In this paper,mathematically interpretable and visual analysis of the convolutional neural network model is performed to obtain the network structure and optimal parameter selection that affect the model performance.This method achieves a significant improvement in the accuracy of the model for image classification and the generalization ability of the model.It has certain practical application value.Simultaneously,a visualization method is used to demonstrate the network structure,parameter changes and model convolutional layer learning process,providing theoretical support for optimizing and understanding deep learning models in practical applications.