Research On Adaptive Selection Of Distance Metric Functions In Semi-Supervised Classification

With the rapid development of computing power and deep learning,deep learning has made leaps and bounds in the fields of images,speech,machine translation,etc.Training a good deep learning model often requires huge sample sizes.However,in actual situations,the cost of obtaining a large amount of labeled data is very high.In some areas such as medical and military,a large amount of labeled data is not available.In contrast,unlabeled data is often easy to obtain.So under this background,the combination of deep learning and semi-supervised learning has become a cutting-edge topic.Prior to this,due to the scarcity of dimension problems and the scarcity of labeled data in traditional supervised learning,semi-supervised learning has been an important research problem in statistical machine learning,and has gradually formed a relatively independent direction,different from supervised learning and unsupervised learning.This article introduces the hypothesis of semi-supervised learning and the classic algorithms of traditional semi-supervised learning: Gaussian mixture model,Semi-Supervised Support Vector Machine,Collaborative training,Lap-based Semi-Supervised Algorithm,etc.For Lap-based semi-supervised model: Lap-SVM.We verified the performance of the model is sensitive to the parameters of the distance measurement function.In response to this problem,this paper proposes a two-stage iterative algorithm to adaptive select the distance measure function when building a Lap(Laplacian matrix)in Lap-SVM.It is further pointed out that this solve process of two-stage iterative algorithm is independent with the selection of distance measure functions.Therefore,the distance measure function can be extended to further optimize the selection of the distance measure function.In numerical experiments,the improved algorithm works on the artificial data set and its performance has improved on actual datasets.More generally,this paper considers learning distance measure function using neural networks.In Chapter 4,we first introduce several algorithms that combination deep learning with semi-supervised learning in recent years.In this paper,the best model of the current regularization constraint class Mean-Teacher is used as a framework,a new network branch is designed to measure the distance between images,and the network is named a similar network.During the training process,theMean-Teacher model being interactive trained with the similar network.On the open datasets SVHN,CIFAR-10,the model's efficiency has been significantly improved.Finally,since the input of similar networks is weakly labeled data,this paper verified that increasing the number of weakly labeled data while maintaining the same labeled data,the model performance can still be improved.Therefore,similar networks have larger applications value.