Visualization Research Of Deep Learning Model Training

In recent years,since the emergence of deep neural network(DNN)as a shining technology,it has played an irreplaceable role in promoting technological innovation.However,deep learning promotes technological progress.At the same time,the decisions made by deep learning models lack explanation and lack of supervision and control over their internal training process.Therefore,it is necessary to understand its internal operating mechanism and understand and analyze the training process of deep learning models by using visualization techniques and visual analysis methods.First,according to the deep learning theory,under the guidance of deep learning experts,and based on the characteristics of the original data,a visualization method for deep learning model training based on the Tensor Flow framework is proposed.In order to help users better understand the entire training process of the neural network architecture,This paper presents a visual analysis design,used to accurately understand,analyze and improve deep neural networks.Secondly,based on the performance analysis of classical convolutional neural networks(CNN),a visual analysis method for explaining deep learning models CNNs is proposed.In order to explain the training process of the deep learning model,based on the CNN network structure,this paper introduces a visual analysis method of multi-layer fusion analysis for the structural characteristics of CNNs.It mainly focuses on the convolution layer,activation layer,pooling layer,fully connected layer and The Softmax layer performs interpretability studies.Combining the requirements of experts for multi-faceted comprehensive diagnosis of CNNs,and based on Alex Net,two real-time visual analysis experiments including convolutional feature maps and activation feature maps are output based on the results learned by the convolutional neural network.Thirdly,on the basis of visual analysis of classical convolutional neural networks,a visualization method for interpreting recurrent neural network models is proposed.The deep learning model is based on the RNN network structure and introduces a visualization method of joint analysis of the network structure for the characteristics of the cyclic structure of the RNN.It mainly includes the visualization methods of the fully connected layer,the cyclic structure,the convolution layer and the pooling layer.Combined with the requirements of RNN comprehensive diagnosis and analysis in the field of natural language processing,based on a simple RNN,a visual analysis experiment including the structure of each component of the recurrent neural network is designed.Finally,under the Ubuntu16.04 operating system,using Python2.7 language under the Tensor Flow1.0 framework,respectively carried out the visual analysis experiment of training CNNs on the MNIST training data set and based on the RNN on the natural language data set to verify this article The validity and versatility of the experiment.