Abstract: In order to improve the classification performance of semi-supervised deep generation models, a new model based on ladder network and improved tri-training algorithm is proposed. This model adds three classifiers to the highest layer of the noisy coding layer of the ladder network, and improves the image classification performance by combining the improved tri-training. First, the labeled data is divided into three parts by using a class-based sampling method. The model adjusts the parameters by combining the labeled error of labeled data and the reconstruction error of unlabeled data, and is trained to get three Large-margin Softmax classifiers. Next, the improved tri-training algorithm is developed to add pseudo-labels to the unlabeled data, and different weights are assigned to the new labeled data to expand the training set. Finally, the expanded training set is applied to update the model. After the above practice, a weighted voting is performed on the classifier to obtain the classification result. The characteristics of the ladder network obtained by this model have better low-dimensional manifold representations, which can effectively avoid classification error caused by uneven distribution of sample data and enhance generalization ability. The model is tested on the MNIST, SVHN and CIFAR10 respectively. In comparison with other semi-supervised deep generation models, test results show that the model proposed in this paper has obtained state-of-the-art classification accuracy over existing semi-supervised learning methods.