Optimization of Generalization Problem based on Mean Teacher Model

This paper focuses on optimizing the Mean Teacher model and validating the improved classification performance on the benchmark CIFAR-10 dataset. The paper optimizes the loss function by constructing a teacher graph, incorporating consistency loss, and introducing smooth neighbors based on the teacher graph. Simultaneously, the student and teacher networks of the original model are replaced with autoencoders to enhance prediction accuracy through the encoder's classification and reconstruction abilities. Ultimately, two ConvLarge structure algorithms, SNTG (Smooth Neighbors on Teacher) and HybridNet, are developed. These three algorithms are compared for recognition performance on the CIFAR-10 dataset, achieving promising results. Both SNTG and HybridNet significantly improve model accuracy compared to the original Mean Teacher algorithm, reducing recognition error rates to around 17% and increasing the accuracy by 3.5%.