Research On Semi-supervised Classification Model Based On Anti-autoencoder

The accuracy of supervised learning depends on the number of labeled samples greatly.The cost of manual labeling is very high.Therefore,semi-supervised learning that uses a large number of unlabeled samples and a limited number of labeled samples has become an effective method of improving the accuracy of the algorithm.The generative model of deep learning that only uses input samples for supervision can help learners use a large number of unlabeled sample data to learn the distribution of samples in semi-supervised learning,which has become a new method in the field of semi-supervised learning.This thesis studies the model structure and training process of the Adversarial Auto-Encoder(AAE)and finds that the model has two different discriminators in semi-supervised classification tasks.These two constraints will weaken the regularization ability of each other during the training process;at the same time,the style of samples generated by the decoder is not cared about in the semi-supervised classification tasks.Experiment proves that the constraint on label variables can also make the distribution of hidden variables close to Gaussian distribution.Therefore,this thesis optimizes the AAE model and proposes the Semi-Supervised Adversarial Auto-Encoder(SSAAE)model.The SSAAE model only has regularization constraints on the label variables,which eliminates the influence of the hidden variable discriminator and the label variable discriminator of the AAE model on regularization.Experiment has been carried out in the MNIST dataset,SVHN dataset and medical images.The SSAAE model has better classification results than the AAE model.Moreover,the optimized model is easier to train and its convergence speed is faster.The AAE model uses a multilayer perceptron as its network infrastructure.Considering that the convolutional neural network has more powerful feature learning abilities than the multilayer perceptron,the multilayer perceptron in the adversarial auto-encoder is replaced by the convolutional neural network structure with stronger feature learning abilities and more stable performance.The thesis proposes Deep Convolutional Adversarial Auto-Encoder(DCAAE)model,which has similar generative adversarial part to the DCGAN model.Experiment shows that the DCAAE model is superior to the SSAAE model in classification accuracy and convergence speed.