Research On Semi-supervised Classification Algorithm Based On Integrated Neural Network

In recent years,the application of deep learning technology in semi-supervised learning(SSL)has set off a boom.Semi-supervised learning is a model training and boosting algorithm combining supervised information samples and unsupervised information samples.This algorithm is more suitable for practical application scenarios.In these scenarios,unlabeled data is easy to obtain and available at any time,while the marking of data is expensive and time-costive.SSL can make full use of unlabeled data to construct a better classifier,which is suitable for the situation of insufficient labeled training data.This paper mainly studies the application of deep learning and ensemble learning in semi-supervised learning:(1)Aiming at the problem that the existing semi-supervised algorithm cannot improve the generalization ability of neural network model in the case of small samples,first,on the standard semi-supervised data set of CIFAR-10,three groups of small sample are trained by using the neural network training.Through the analysis of the experimental results,a semi-supervised learning algorithm integrated with neural network is proposed.The algorithm improves the performance of single model by simplifying the learning task of a single network,and effectively combines the output results of multiple single models to complete the sample screening process.The algorithm is divided into training stage and marking stage.In the training stage,a three-classification model is trained between each two categories.The third category is the background class that is randomly sampled(The background class is obtained by sampling other samples that do not belong to these two categories).According to the classification categories of the model,the output results are combined into a category filter;in the marking stage,the category filter weights the output according to the performance of the model in the verification set to obtain the class confidence of the samples,the samples whose confidence level are mutually exclusive are filtered out.In order to improve the performance of the network,a new sample is generated by weighted fusion of the reserved samples and the labeled samples of the same category.Compared with other semi supervised algorithms,the experimental results of the proposed algorithm are improved on different semi supervised learning test conditions,and the error rate is reduced by 1% ? 2%.(2)Aiming at the problem that Temporal ensembling algorithm based on consistency loss is difficult to implement in practical application scenarios,an improved scheme is proposed.It mainly includes two aspects of improvement: the original algorithm mainly relies on a small number of supervised samples in the initial training stage,and the number of effective samples becomes an important problem in the restricted model.In this paper,a self-supervised pre-training method based on the rotation angle prediction of unsupervised information samples is proposed.This method is used to train and transfer network weights to subsequent tasks on unsupervised samples,which improves the performance of the network at the initial stage of training.Aiming at the disadvantage of the original algorithm's loss function that it cannot accurately track the current performance of the network,an improved calculation method for the weight value of the loss function is proposed,and the average gradient value of the cross-entropy loss term in the period range is introduced as the weight value of the consistency loss.On the supervised dataset with small sample size,the classification effect of the improved algorithm is much better than that of the original algorithm and the classification error rate is reduced by 2% ? 20%.(3)Finally,the main research content of this paper is summarized.The future work of the research content that need to be further improved is pointed out.