Research And Application Of Convolutional Neural Network In Collaborative Semi-Supervised Classification

Classification as an important data processing method in machine learning is widely used in many fields.Among them,the problem of image classification has always been a hot and difficult point of research.An effective way to solve this difficulty is convolutional neural networks However,training a convolutional neural network with excellent performance requires the support of a large amount of labeled data.This is an extremely heavy task,so the semi-supervised learning method that uses a small amount of labeled data and a large amount of unlabeled data is more research value and practical significanceIn the semi-supervised classification of convolutional neural networks,the collaborative methods have always attracted the attention of many scholars.In the collaborative method,by copying the neural network model into two copies,one as the teacher model and the other as the student model,forcing the two models to be consistent on the label prediction of the unlabeled data to use the information carried by the unlabeled data which has a good classification effect.However,this method still has some shortcomings:1.The degree of coupling between the two models in the collaborative method is too high,which leads to a performance bottleneck in the method itself;2.Only relying on two student models cannot effectively explore the entire solution space;3.When the number of student model increases,the single teacher model cannot effectively guide the student models.In order to solve the above three problems,this paper focuses on the collaborative semi-supervised classification algorithm and makes some improvements to it.The main research work and results are as follows:(1)Aiming at the problem of high coupling between the two models in the collaborative method,a dual-student algorithm based on collaboration is proposed.In the original collaborative algorithm,because the parameter update of the teacher model is based on the student model,there is a high degree of coupling between the two,resulting in the performance bottleneck of the algorithm.In order to overcome this problem,the algorithm participates in the parameter update process of the teacher model by adding a new and independent student model to achieve the purpose of weakening the coupling.At the same time,the classification loss of the two student models is used as the criterion to measure the two students.The performance of the model is fed back to the parameter update process of the teacher model.We apply the algorithm to handwritten digit recognition and natural image classification.The experimental results show that this algorithm can effectively weaken the coupling and improve the classification accuracy.(2)Aiming at the problem that two student models cannot effectively explore the entire solution space,a multiple student algorithm based on group optimization is proposed.Although the dual student algorithm overcomes the strong coupling defect of the collaborative method and expands the solution space,it cannot effectively explore the entire solution space using only two student models.Therefore,we expand the dual students to multiple students,and use the method of group optimization to weigh the contribution of each student model parameter to the teacher model parameter,and then use the designed historical optimal parameters and batch optimal parameters to attach the loose constraints to each student models and ensure the consistency between each student model and the teacher model,so that the entire model can get a convergent solution.We apply this algorithm to natural image data sets such as SVHN and CIFAR-10.The experimental results show that this algorithm can effectively explore the solution space and improve the accuracy of classification.(3)Aiming at the problem that a single teacher model cannot effectively guide the training of student models when the number of student models increases,a multiple student algorithm based on feedforward network is proposed.Although the multi-student algorithm based on group optimization improves the accuracy of image classification,when the number of student models exceeds a certain value,the performance of the algorithm is not obvious.The reason is that the number of teacher models is small,which cannot effectively guide the student models.In order to overcome this problem,we extend the teacher model to multiple,construct a feedforward network by adopting a fully connected way between the student group and the teacher group,and use the classification loss as the criterion to adjust the connection weights to achieve optimal performance.We apply this algorithm to the SVHN,CIFA-10,CIFAR-100 data sets,and the experimental results show that the algorithm can further improve the accuracy of classification.