Neural Networks For Small Sample Data Classification Intergraded With Decentralized Technology

Data classification is a hot research topic.Small sample data classification problem is one of the most popular problems.In some particular areas,the number of labeled samples is limited.Hence,classification algorithms with small sample size are designed.Among different small sample data sets,gene expression data sets are most representatives,which can be classified effectively by using cost-sensitive algorithms.Image classification is also a popular application area for data classification.The number of parameters in the convolutional neural network used for image classification is large.And calculating these parameters requires tremendous computational power.Therefore,this dissertation proposes improved models and a distributed computing architecture based on decentralized technology.Following studies were mainly conducted in this dissertation:1)Firstly,unbalanced data set classification problem is studied.In this dissertation,the cost sensitive method is used to improve the classification of unbalanced data sets.Accurate determination of the cost weights is completed through grid search and function fitting.The method proposed in this dissertation can provide better Weighted Classification Accuracy(WCA)than the manual determination method and genetic algorithm.Compared with the traditional accuracy calculation method,WCA is more representative in the classification of unbalanced data sets.Experimental results show that the grid search method can preliminarily complete the classification of unbalanced data sets.The function fitting method can find the optimal weights more accurately.2)The second part of this dissertation proposes a convolution neural network for small sample image classification,named random convolutional network.Adopting the idea of random network weights from ELM,combining with Inception structure,a classification network is designed for small data sets,which can be rapidly trained and avoid overfitting.Based on the model structure,the back propagation of the network is limited to reduce the training time.Meanwhile,increasing multi-scale network branches can improve classification accuracy.The newly proposed random convolution neural network has to tune many parameters,which is hardly done by empirical determination.With design of blockchain architecture,a decentralized computing network is proposed,in which each "miner" finds the optimal solution through competition.Experimental results show that in small sample data sets,random convolutional network has better classification performance,while traditional convolutional neural networks always fail to obtain good classification results.