Semi-supervised Image Classification Based On Broad Network

In real world,it’s difficult to obtain labeled samples,which leads to low generalization ability of models.Semi-supervised learning mainly studies how to use a large number of unlabeled samples to improve the performance of the model when there are few labeled samples.At the same time,broad network has flexible structure and efficient learning process.So,this thesis mainly focuses on the semi-supervised image classification based on broad network.The main research results and contributions are as follows:The linear sparse mapping method in broad network is difficult to effectively characterize the complex nonlinear features of data and the lack of labeled samples leads to poor generalization performance of broad network.Aiming at solve these two problems,we propose an autoencoder and hypergraph-based semi-supervised broad network.First we use all samples including labeled samples and unlabeled samples to train the autoencoder,and the trained autoencoder is used to automatically extract the complex nonlinear features of the data;Second,the feature layers of auto encoder are seen as the feature nodes and then generate the enhance nodes;Third,we constructed a hypergraph of data to describe the higher-order manifold relationship between labeled samples and unlabeled samples,and the hypergraph is introduced into the loss function of our model;Finally,we use ridge regression method to solve the loss function of our model.The graph-based semi-supervised algorithm have a Laplace matrix which is difficult to calculate.The generative adversarial networks are hard to apply in semisupervised learning.Aiming at solve these two problems,we purpose a generative adversarial semi-supervised broad network.First,we use all samples to train the variational autoencoder and then use this trained variational autoencoder to generate pseudo samples;Second,we take the pseudo samples as a new class samples,and use these samples to train the convolutional broad network;Third,we make the variational autoencoder and the convolutional broad network to train each other by redesign their loss functions.Finally,our model can achieve better semi-supervised image classification results.As we all know,there is a large deviation between the distribution of labeled samples and real-world samples in the semi-supervised learning.We call the name of this problem is domain shift.So,we purpose a domain expansion and residual connection based broad network.First,we propose an inductive bias-adaptive sample augmentation method to expand the domain of labeled samples;Second,we use expanded samples to train our residual connection based broad network;Third,we use the unlabeled samples which have a high confidence by pseudo labeling them in the training process until our model is convergence.Experimental results are performed on the MNIST,NORB and SVHN datasets show that our methods in this thesis can achieve the high semi-supervised classification accuracy.