Research On Neural Network Architecture Of Component Adaptive

With the advent of the information age,artificial intelligence has changed from academic research to application-driven.Intelligent systems are used for cognition,recognition,analysis,and decision-making.The essence and ultimate goal are to simulate the process of human consciousness and thinking.Due to a large amount of data,complex deep non-linear model and computational performance,the current technology development of deep learning is created,a new wave of artificial intelligence is set off.Therefore,data collection,algorithm design,efficient calculation,and other technologies are the key core roles in intelligent development.Some components play significant roles in various neural network models,such as convolution and pooling in convolutional neural networks,feature fusion components in inception networks.How to perform effective hyperparameter,model design,and search in the possible space is the fundamental problems for the entire architecture to be efficiently trained and applied.Through related basic research on components,different data scenarios require different network component hyperparameter settings,and a large amount of data is a condition for training intelligent models.Any key technologies and models need to be supported by the data.Then it can be applied in different scenarios so that those intelligent technologies can be used and developed on a large scale industry.The network model scale design is getting larger and deeper.There are more and more network nodes,makes the training process requires a lot of resources.Therefore,how to effectively release the burden of computing resources and ensure the certain performance of the model is the research focus of the future.According to the data and model problems existing in the neural network,we from network components,data applications and shallow architectures respectively to deal with the main challenges.With the research ideas of "network components-scene analysis-architecture stretch",we focus on the following three research points to carry out the work:(1)Adaptive select the convolution kernel components in convolutional networks,analysis the feature fusion components and identity mapping components in shallow networks,propose Tception models;(2)We design a Tception CAE model based on medical text question & answer and make some cluster application,explore the semantic relationship and topic analysis between the different clustering in experiments;(3)According to the metabolism mechanism of biological cells,combine with the problem of node resource consumption in neural networks,parameter learning through the addition of hidden layer neurons and autophagy process,so that the model architecture automatically performs extended learning to form an architecture which dynamic stretch network on the broad direction.For the neural network,how to adaptively set hyperparameters and components is a fundamental problem.We focus on convolutional neural networks,as multi-convolution kernel settings need some prior knowledge and different feature fusion methods in different scenarios are very important.According to the adaptive selection of convolution kernel,feature fusion and identity mapping method analysis,the Tception model is proposed.With single convolution kernel,feature extraction is limited.The multiple convolution kernels can acquire complex features that capture the spatial correlation between information elements,we used a multi-convolution kernel adaptive select method based on ensemble learning theory to increase feature diversity.Features fusion can distinguish multiple features in the data from different levels with a centralized manner,eliminate redundancy between those features,we proposed different feature fusion methods from the relationship between features.The feature reuse can improve model performance and propose different identity mapping enhances feature propagation,encourages feature reuse,analyzes and explores residuals in shallow networks.Then,how to analyze data characteristics and select components in different scenarios is a practical problem in the application process.We focus on medical text question & answer data.We propose a Tception Convolutional Autoencoder model(Tception CAE)used in medical text clustering application for those challenges.We analysis user medical question and answer text data which was collected from the medical platform.The user's description of the disease has some problems such as lack of professionalism,sparse text,high-dimensional semantics,data is not easy to be labelled,the deviation of label information,etc.We use a convolution autoencoder model for unsupervised feature representation learning in unlabeled data.Selection of convolution kernels according to practical application tasks with clustering ensemble theory.Consider the basic network component design from the perspective of data characteristics,and conduct various experimental comparisons of the methods proposed in the network component through real text data.Obtain the clustering results of the semantic relationship,subject word cloud,and Q & A correlation analysis.At last,how to effectively release the computational resource burden of neural network models,but it can ensure a certain performance of the model is a key issue for future deep learning research.Inspired by the proliferation and autophagy of cell metabolism,we focus on the self-adjusting mechanism of the number of neurons per layer.Aiming at how to carry out network metabolism through the proliferation of new neurons and phagocytosis of decay neurons,a dynamic stretch network on the broad direction(DSN)with universal applicability is proposed from the perspective of network neurons.DSN uses the dynamic growth of the network that scientifically and comprehensively to explore the extended behaviour of shallow neural networks.First initialize a given small network model,according to the new conditions for neurons proposed in DSN.Static or dynamic techniques are used to generate new hidden neurons,and different approaches to learn the weights of these neurons based on metalearning.In order to maintain the efficiency of the model,we combined with the autophagy conditions in the network,the neurons that need to be deleted are selected for elimination before the network needs to be updated,study the self-adjusting function of the network model to maximize the resources and efficiency.