Semi-supervised Classification Based On GAN Minds

For classification problems,the research of supervised learning which based on a large amount of labeled data has achieved a lot of results.But in reality,it is very difficult to obtain sufficient labeled samples.In fact,most of the data we can get is unlabeled or the tag values of the data are mostly missing.Therefore,the semi-supervised learning algorithm which can help improve the generalization ability of the model by learning the data information in large amounts of unlabeled data has always been the hot and difficult point in our research.The common algorithms for semi-supervised learning can be divided into low-density separation methods,graph-based methods,disagreement methods and generative methods,in which the generative adversarial network in the method is the object of this paper.As a hot spot of current machine learning research,the generative adversarial network(GAN)uses the mind of zero-sum game.The discriminator and generator make up the GAN model,this two elements are opposed to each other in order to optimizing their own model.The output of the discriminator is the probability assigned to the input data,and the generator learns the distribution of real data.The discriminator consist of the neural networks updates the parameters by ascending its stochastic gradient about the losses of assigning the the low probability value to the input which belong to real data and assigning the high probability value to the input which belong to generated data.Similarly,the generator consist of the neural networks updates the parameters according to the losses formed by the discriminator assigned the low probability value to the input about generated data.The discriminator and generator optimize alternately,finally reaching the Nash equilibrium,that is a state about the discriminator can't distinguish the input data is belong to real data or generated data.In this paper,we use Improved GAN's semi-supervised classification mind,add a weak classifier which trained with labeled data in front of GAN model.And the prediction results of the unlabeled samples by the weak classifier are regarded as the pseudo-labeled value of the sample.Then,we modify the loss function of the original model and use these small amount of labeled data and the original unlabeled data with "fake" labels to train GAN model.After the condition of updating the label of unlabeled data set is satisfied,we update the pseudo-label of the sample that exceeds a function threshold.After several times of updating,the label value of the sample is basically close to the true label of the unlabeled data.This algorithm is applied to three common image data set for validation,we found that the algorithm can mitigate the problem that the classification accuracy of the semi-supervised classification model with a few of labeled data is low.At the same time,GAN's generator can also learn the distribution of real data,generate samples with reasonable authenticity and clarity,and it's almost no mode collapse phenomenon.